Loading...
1// SPDX-License-Identifier: GPL-2.0
2
3#include <linux/bitops.h>
4#include <linux/slab.h>
5#include <linux/bio.h>
6#include <linux/mm.h>
7#include <linux/pagemap.h>
8#include <linux/page-flags.h>
9#include <linux/sched/mm.h>
10#include <linux/spinlock.h>
11#include <linux/blkdev.h>
12#include <linux/swap.h>
13#include <linux/writeback.h>
14#include <linux/pagevec.h>
15#include <linux/prefetch.h>
16#include <linux/fsverity.h>
17#include "misc.h"
18#include "extent_io.h"
19#include "extent-io-tree.h"
20#include "extent_map.h"
21#include "ctree.h"
22#include "btrfs_inode.h"
23#include "bio.h"
24#include "locking.h"
25#include "rcu-string.h"
26#include "backref.h"
27#include "disk-io.h"
28#include "subpage.h"
29#include "zoned.h"
30#include "block-group.h"
31#include "compression.h"
32#include "fs.h"
33#include "accessors.h"
34#include "file-item.h"
35#include "file.h"
36#include "dev-replace.h"
37#include "super.h"
38#include "transaction.h"
39
40static struct kmem_cache *extent_buffer_cache;
41
42#ifdef CONFIG_BTRFS_DEBUG
43static inline void btrfs_leak_debug_add_eb(struct extent_buffer *eb)
44{
45 struct btrfs_fs_info *fs_info = eb->fs_info;
46 unsigned long flags;
47
48 spin_lock_irqsave(&fs_info->eb_leak_lock, flags);
49 list_add(&eb->leak_list, &fs_info->allocated_ebs);
50 spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
51}
52
53static inline void btrfs_leak_debug_del_eb(struct extent_buffer *eb)
54{
55 struct btrfs_fs_info *fs_info = eb->fs_info;
56 unsigned long flags;
57
58 spin_lock_irqsave(&fs_info->eb_leak_lock, flags);
59 list_del(&eb->leak_list);
60 spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
61}
62
63void btrfs_extent_buffer_leak_debug_check(struct btrfs_fs_info *fs_info)
64{
65 struct extent_buffer *eb;
66 unsigned long flags;
67
68 /*
69 * If we didn't get into open_ctree our allocated_ebs will not be
70 * initialized, so just skip this.
71 */
72 if (!fs_info->allocated_ebs.next)
73 return;
74
75 WARN_ON(!list_empty(&fs_info->allocated_ebs));
76 spin_lock_irqsave(&fs_info->eb_leak_lock, flags);
77 while (!list_empty(&fs_info->allocated_ebs)) {
78 eb = list_first_entry(&fs_info->allocated_ebs,
79 struct extent_buffer, leak_list);
80 pr_err(
81 "BTRFS: buffer leak start %llu len %lu refs %d bflags %lu owner %llu\n",
82 eb->start, eb->len, atomic_read(&eb->refs), eb->bflags,
83 btrfs_header_owner(eb));
84 list_del(&eb->leak_list);
85 kmem_cache_free(extent_buffer_cache, eb);
86 }
87 spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
88}
89#else
90#define btrfs_leak_debug_add_eb(eb) do {} while (0)
91#define btrfs_leak_debug_del_eb(eb) do {} while (0)
92#endif
93
94/*
95 * Structure to record info about the bio being assembled, and other info like
96 * how many bytes are there before stripe/ordered extent boundary.
97 */
98struct btrfs_bio_ctrl {
99 struct btrfs_bio *bbio;
100 enum btrfs_compression_type compress_type;
101 u32 len_to_oe_boundary;
102 blk_opf_t opf;
103 btrfs_bio_end_io_t end_io_func;
104 struct writeback_control *wbc;
105};
106
107static void submit_one_bio(struct btrfs_bio_ctrl *bio_ctrl)
108{
109 struct btrfs_bio *bbio = bio_ctrl->bbio;
110
111 if (!bbio)
112 return;
113
114 /* Caller should ensure the bio has at least some range added */
115 ASSERT(bbio->bio.bi_iter.bi_size);
116
117 if (btrfs_op(&bbio->bio) == BTRFS_MAP_READ &&
118 bio_ctrl->compress_type != BTRFS_COMPRESS_NONE)
119 btrfs_submit_compressed_read(bbio);
120 else
121 btrfs_submit_bio(bbio, 0);
122
123 /* The bbio is owned by the end_io handler now */
124 bio_ctrl->bbio = NULL;
125}
126
127/*
128 * Submit or fail the current bio in the bio_ctrl structure.
129 */
130static void submit_write_bio(struct btrfs_bio_ctrl *bio_ctrl, int ret)
131{
132 struct btrfs_bio *bbio = bio_ctrl->bbio;
133
134 if (!bbio)
135 return;
136
137 if (ret) {
138 ASSERT(ret < 0);
139 btrfs_bio_end_io(bbio, errno_to_blk_status(ret));
140 /* The bio is owned by the end_io handler now */
141 bio_ctrl->bbio = NULL;
142 } else {
143 submit_one_bio(bio_ctrl);
144 }
145}
146
147int __init extent_buffer_init_cachep(void)
148{
149 extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
150 sizeof(struct extent_buffer), 0,
151 SLAB_MEM_SPREAD, NULL);
152 if (!extent_buffer_cache)
153 return -ENOMEM;
154
155 return 0;
156}
157
158void __cold extent_buffer_free_cachep(void)
159{
160 /*
161 * Make sure all delayed rcu free are flushed before we
162 * destroy caches.
163 */
164 rcu_barrier();
165 kmem_cache_destroy(extent_buffer_cache);
166}
167
168void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
169{
170 unsigned long index = start >> PAGE_SHIFT;
171 unsigned long end_index = end >> PAGE_SHIFT;
172 struct page *page;
173
174 while (index <= end_index) {
175 page = find_get_page(inode->i_mapping, index);
176 BUG_ON(!page); /* Pages should be in the extent_io_tree */
177 clear_page_dirty_for_io(page);
178 put_page(page);
179 index++;
180 }
181}
182
183static void process_one_page(struct btrfs_fs_info *fs_info,
184 struct page *page, struct page *locked_page,
185 unsigned long page_ops, u64 start, u64 end)
186{
187 struct folio *folio = page_folio(page);
188 u32 len;
189
190 ASSERT(end + 1 - start != 0 && end + 1 - start < U32_MAX);
191 len = end + 1 - start;
192
193 if (page_ops & PAGE_SET_ORDERED)
194 btrfs_folio_clamp_set_ordered(fs_info, folio, start, len);
195 if (page_ops & PAGE_START_WRITEBACK) {
196 btrfs_folio_clamp_clear_dirty(fs_info, folio, start, len);
197 btrfs_folio_clamp_set_writeback(fs_info, folio, start, len);
198 }
199 if (page_ops & PAGE_END_WRITEBACK)
200 btrfs_folio_clamp_clear_writeback(fs_info, folio, start, len);
201
202 if (page != locked_page && (page_ops & PAGE_UNLOCK))
203 btrfs_folio_end_writer_lock(fs_info, folio, start, len);
204}
205
206static void __process_pages_contig(struct address_space *mapping,
207 struct page *locked_page, u64 start, u64 end,
208 unsigned long page_ops)
209{
210 struct btrfs_fs_info *fs_info = btrfs_sb(mapping->host->i_sb);
211 pgoff_t start_index = start >> PAGE_SHIFT;
212 pgoff_t end_index = end >> PAGE_SHIFT;
213 pgoff_t index = start_index;
214 struct folio_batch fbatch;
215 int i;
216
217 folio_batch_init(&fbatch);
218 while (index <= end_index) {
219 int found_folios;
220
221 found_folios = filemap_get_folios_contig(mapping, &index,
222 end_index, &fbatch);
223 for (i = 0; i < found_folios; i++) {
224 struct folio *folio = fbatch.folios[i];
225
226 process_one_page(fs_info, &folio->page, locked_page,
227 page_ops, start, end);
228 }
229 folio_batch_release(&fbatch);
230 cond_resched();
231 }
232}
233
234static noinline void __unlock_for_delalloc(struct inode *inode,
235 struct page *locked_page,
236 u64 start, u64 end)
237{
238 unsigned long index = start >> PAGE_SHIFT;
239 unsigned long end_index = end >> PAGE_SHIFT;
240
241 ASSERT(locked_page);
242 if (index == locked_page->index && end_index == index)
243 return;
244
245 __process_pages_contig(inode->i_mapping, locked_page, start, end,
246 PAGE_UNLOCK);
247}
248
249static noinline int lock_delalloc_pages(struct inode *inode,
250 struct page *locked_page,
251 u64 start,
252 u64 end)
253{
254 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
255 struct address_space *mapping = inode->i_mapping;
256 pgoff_t start_index = start >> PAGE_SHIFT;
257 pgoff_t end_index = end >> PAGE_SHIFT;
258 pgoff_t index = start_index;
259 u64 processed_end = start;
260 struct folio_batch fbatch;
261
262 if (index == locked_page->index && index == end_index)
263 return 0;
264
265 folio_batch_init(&fbatch);
266 while (index <= end_index) {
267 unsigned int found_folios, i;
268
269 found_folios = filemap_get_folios_contig(mapping, &index,
270 end_index, &fbatch);
271 if (found_folios == 0)
272 goto out;
273
274 for (i = 0; i < found_folios; i++) {
275 struct folio *folio = fbatch.folios[i];
276 struct page *page = folio_page(folio, 0);
277 u32 len = end + 1 - start;
278
279 if (page == locked_page)
280 continue;
281
282 if (btrfs_folio_start_writer_lock(fs_info, folio, start,
283 len))
284 goto out;
285
286 if (!PageDirty(page) || page->mapping != mapping) {
287 btrfs_folio_end_writer_lock(fs_info, folio, start,
288 len);
289 goto out;
290 }
291
292 processed_end = page_offset(page) + PAGE_SIZE - 1;
293 }
294 folio_batch_release(&fbatch);
295 cond_resched();
296 }
297
298 return 0;
299out:
300 folio_batch_release(&fbatch);
301 if (processed_end > start)
302 __unlock_for_delalloc(inode, locked_page, start, processed_end);
303 return -EAGAIN;
304}
305
306/*
307 * Find and lock a contiguous range of bytes in the file marked as delalloc, no
308 * more than @max_bytes.
309 *
310 * @start: The original start bytenr to search.
311 * Will store the extent range start bytenr.
312 * @end: The original end bytenr of the search range
313 * Will store the extent range end bytenr.
314 *
315 * Return true if we find a delalloc range which starts inside the original
316 * range, and @start/@end will store the delalloc range start/end.
317 *
318 * Return false if we can't find any delalloc range which starts inside the
319 * original range, and @start/@end will be the non-delalloc range start/end.
320 */
321EXPORT_FOR_TESTS
322noinline_for_stack bool find_lock_delalloc_range(struct inode *inode,
323 struct page *locked_page, u64 *start,
324 u64 *end)
325{
326 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
327 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
328 const u64 orig_start = *start;
329 const u64 orig_end = *end;
330 /* The sanity tests may not set a valid fs_info. */
331 u64 max_bytes = fs_info ? fs_info->max_extent_size : BTRFS_MAX_EXTENT_SIZE;
332 u64 delalloc_start;
333 u64 delalloc_end;
334 bool found;
335 struct extent_state *cached_state = NULL;
336 int ret;
337 int loops = 0;
338
339 /* Caller should pass a valid @end to indicate the search range end */
340 ASSERT(orig_end > orig_start);
341
342 /* The range should at least cover part of the page */
343 ASSERT(!(orig_start >= page_offset(locked_page) + PAGE_SIZE ||
344 orig_end <= page_offset(locked_page)));
345again:
346 /* step one, find a bunch of delalloc bytes starting at start */
347 delalloc_start = *start;
348 delalloc_end = 0;
349 found = btrfs_find_delalloc_range(tree, &delalloc_start, &delalloc_end,
350 max_bytes, &cached_state);
351 if (!found || delalloc_end <= *start || delalloc_start > orig_end) {
352 *start = delalloc_start;
353
354 /* @delalloc_end can be -1, never go beyond @orig_end */
355 *end = min(delalloc_end, orig_end);
356 free_extent_state(cached_state);
357 return false;
358 }
359
360 /*
361 * start comes from the offset of locked_page. We have to lock
362 * pages in order, so we can't process delalloc bytes before
363 * locked_page
364 */
365 if (delalloc_start < *start)
366 delalloc_start = *start;
367
368 /*
369 * make sure to limit the number of pages we try to lock down
370 */
371 if (delalloc_end + 1 - delalloc_start > max_bytes)
372 delalloc_end = delalloc_start + max_bytes - 1;
373
374 /* step two, lock all the pages after the page that has start */
375 ret = lock_delalloc_pages(inode, locked_page,
376 delalloc_start, delalloc_end);
377 ASSERT(!ret || ret == -EAGAIN);
378 if (ret == -EAGAIN) {
379 /* some of the pages are gone, lets avoid looping by
380 * shortening the size of the delalloc range we're searching
381 */
382 free_extent_state(cached_state);
383 cached_state = NULL;
384 if (!loops) {
385 max_bytes = PAGE_SIZE;
386 loops = 1;
387 goto again;
388 } else {
389 found = false;
390 goto out_failed;
391 }
392 }
393
394 /* step three, lock the state bits for the whole range */
395 lock_extent(tree, delalloc_start, delalloc_end, &cached_state);
396
397 /* then test to make sure it is all still delalloc */
398 ret = test_range_bit(tree, delalloc_start, delalloc_end,
399 EXTENT_DELALLOC, cached_state);
400 if (!ret) {
401 unlock_extent(tree, delalloc_start, delalloc_end,
402 &cached_state);
403 __unlock_for_delalloc(inode, locked_page,
404 delalloc_start, delalloc_end);
405 cond_resched();
406 goto again;
407 }
408 free_extent_state(cached_state);
409 *start = delalloc_start;
410 *end = delalloc_end;
411out_failed:
412 return found;
413}
414
415void extent_clear_unlock_delalloc(struct btrfs_inode *inode, u64 start, u64 end,
416 struct page *locked_page,
417 u32 clear_bits, unsigned long page_ops)
418{
419 clear_extent_bit(&inode->io_tree, start, end, clear_bits, NULL);
420
421 __process_pages_contig(inode->vfs_inode.i_mapping, locked_page,
422 start, end, page_ops);
423}
424
425static bool btrfs_verify_page(struct page *page, u64 start)
426{
427 if (!fsverity_active(page->mapping->host) ||
428 PageUptodate(page) ||
429 start >= i_size_read(page->mapping->host))
430 return true;
431 return fsverity_verify_page(page);
432}
433
434static void end_page_read(struct page *page, bool uptodate, u64 start, u32 len)
435{
436 struct btrfs_fs_info *fs_info = btrfs_sb(page->mapping->host->i_sb);
437 struct folio *folio = page_folio(page);
438
439 ASSERT(page_offset(page) <= start &&
440 start + len <= page_offset(page) + PAGE_SIZE);
441
442 if (uptodate && btrfs_verify_page(page, start))
443 btrfs_folio_set_uptodate(fs_info, folio, start, len);
444 else
445 btrfs_folio_clear_uptodate(fs_info, folio, start, len);
446
447 if (!btrfs_is_subpage(fs_info, page->mapping))
448 unlock_page(page);
449 else
450 btrfs_subpage_end_reader(fs_info, folio, start, len);
451}
452
453/*
454 * After a write IO is done, we need to:
455 *
456 * - clear the uptodate bits on error
457 * - clear the writeback bits in the extent tree for the range
458 * - filio_end_writeback() if there is no more pending io for the folio
459 *
460 * Scheduling is not allowed, so the extent state tree is expected
461 * to have one and only one object corresponding to this IO.
462 */
463static void end_bbio_data_write(struct btrfs_bio *bbio)
464{
465 struct bio *bio = &bbio->bio;
466 int error = blk_status_to_errno(bio->bi_status);
467 struct folio_iter fi;
468
469 ASSERT(!bio_flagged(bio, BIO_CLONED));
470 bio_for_each_folio_all(fi, bio) {
471 struct folio *folio = fi.folio;
472 struct inode *inode = folio->mapping->host;
473 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
474 const u32 sectorsize = fs_info->sectorsize;
475 u64 start = folio_pos(folio) + fi.offset;
476 u32 len = fi.length;
477
478 /* Only order 0 (single page) folios are allowed for data. */
479 ASSERT(folio_order(folio) == 0);
480
481 /* Our read/write should always be sector aligned. */
482 if (!IS_ALIGNED(fi.offset, sectorsize))
483 btrfs_err(fs_info,
484 "partial page write in btrfs with offset %zu and length %zu",
485 fi.offset, fi.length);
486 else if (!IS_ALIGNED(fi.length, sectorsize))
487 btrfs_info(fs_info,
488 "incomplete page write with offset %zu and length %zu",
489 fi.offset, fi.length);
490
491 btrfs_finish_ordered_extent(bbio->ordered,
492 folio_page(folio, 0), start, len, !error);
493 if (error)
494 mapping_set_error(folio->mapping, error);
495 btrfs_folio_clear_writeback(fs_info, folio, start, len);
496 }
497
498 bio_put(bio);
499}
500
501/*
502 * Record previously processed extent range
503 *
504 * For endio_readpage_release_extent() to handle a full extent range, reducing
505 * the extent io operations.
506 */
507struct processed_extent {
508 struct btrfs_inode *inode;
509 /* Start of the range in @inode */
510 u64 start;
511 /* End of the range in @inode */
512 u64 end;
513 bool uptodate;
514};
515
516/*
517 * Try to release processed extent range
518 *
519 * May not release the extent range right now if the current range is
520 * contiguous to processed extent.
521 *
522 * Will release processed extent when any of @inode, @uptodate, the range is
523 * no longer contiguous to the processed range.
524 *
525 * Passing @inode == NULL will force processed extent to be released.
526 */
527static void endio_readpage_release_extent(struct processed_extent *processed,
528 struct btrfs_inode *inode, u64 start, u64 end,
529 bool uptodate)
530{
531 struct extent_state *cached = NULL;
532 struct extent_io_tree *tree;
533
534 /* The first extent, initialize @processed */
535 if (!processed->inode)
536 goto update;
537
538 /*
539 * Contiguous to processed extent, just uptodate the end.
540 *
541 * Several things to notice:
542 *
543 * - bio can be merged as long as on-disk bytenr is contiguous
544 * This means we can have page belonging to other inodes, thus need to
545 * check if the inode still matches.
546 * - bvec can contain range beyond current page for multi-page bvec
547 * Thus we need to do processed->end + 1 >= start check
548 */
549 if (processed->inode == inode && processed->uptodate == uptodate &&
550 processed->end + 1 >= start && end >= processed->end) {
551 processed->end = end;
552 return;
553 }
554
555 tree = &processed->inode->io_tree;
556 /*
557 * Now we don't have range contiguous to the processed range, release
558 * the processed range now.
559 */
560 unlock_extent(tree, processed->start, processed->end, &cached);
561
562update:
563 /* Update processed to current range */
564 processed->inode = inode;
565 processed->start = start;
566 processed->end = end;
567 processed->uptodate = uptodate;
568}
569
570static void begin_page_read(struct btrfs_fs_info *fs_info, struct page *page)
571{
572 struct folio *folio = page_folio(page);
573
574 ASSERT(folio_test_locked(folio));
575 if (!btrfs_is_subpage(fs_info, folio->mapping))
576 return;
577
578 ASSERT(folio_test_private(folio));
579 btrfs_subpage_start_reader(fs_info, folio, page_offset(page), PAGE_SIZE);
580}
581
582/*
583 * After a data read IO is done, we need to:
584 *
585 * - clear the uptodate bits on error
586 * - set the uptodate bits if things worked
587 * - set the folio up to date if all extents in the tree are uptodate
588 * - clear the lock bit in the extent tree
589 * - unlock the folio if there are no other extents locked for it
590 *
591 * Scheduling is not allowed, so the extent state tree is expected
592 * to have one and only one object corresponding to this IO.
593 */
594static void end_bbio_data_read(struct btrfs_bio *bbio)
595{
596 struct bio *bio = &bbio->bio;
597 struct processed_extent processed = { 0 };
598 struct folio_iter fi;
599 /*
600 * The offset to the beginning of a bio, since one bio can never be
601 * larger than UINT_MAX, u32 here is enough.
602 */
603 u32 bio_offset = 0;
604
605 ASSERT(!bio_flagged(bio, BIO_CLONED));
606 bio_for_each_folio_all(fi, &bbio->bio) {
607 bool uptodate = !bio->bi_status;
608 struct folio *folio = fi.folio;
609 struct inode *inode = folio->mapping->host;
610 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
611 const u32 sectorsize = fs_info->sectorsize;
612 u64 start;
613 u64 end;
614 u32 len;
615
616 /* For now only order 0 folios are supported for data. */
617 ASSERT(folio_order(folio) == 0);
618 btrfs_debug(fs_info,
619 "%s: bi_sector=%llu, err=%d, mirror=%u",
620 __func__, bio->bi_iter.bi_sector, bio->bi_status,
621 bbio->mirror_num);
622
623 /*
624 * We always issue full-sector reads, but if some block in a
625 * folio fails to read, blk_update_request() will advance
626 * bv_offset and adjust bv_len to compensate. Print a warning
627 * for unaligned offsets, and an error if they don't add up to
628 * a full sector.
629 */
630 if (!IS_ALIGNED(fi.offset, sectorsize))
631 btrfs_err(fs_info,
632 "partial page read in btrfs with offset %zu and length %zu",
633 fi.offset, fi.length);
634 else if (!IS_ALIGNED(fi.offset + fi.length, sectorsize))
635 btrfs_info(fs_info,
636 "incomplete page read with offset %zu and length %zu",
637 fi.offset, fi.length);
638
639 start = folio_pos(folio) + fi.offset;
640 end = start + fi.length - 1;
641 len = fi.length;
642
643 if (likely(uptodate)) {
644 loff_t i_size = i_size_read(inode);
645 pgoff_t end_index = i_size >> folio_shift(folio);
646
647 /*
648 * Zero out the remaining part if this range straddles
649 * i_size.
650 *
651 * Here we should only zero the range inside the folio,
652 * not touch anything else.
653 *
654 * NOTE: i_size is exclusive while end is inclusive.
655 */
656 if (folio_index(folio) == end_index && i_size <= end) {
657 u32 zero_start = max(offset_in_folio(folio, i_size),
658 offset_in_folio(folio, start));
659 u32 zero_len = offset_in_folio(folio, end) + 1 -
660 zero_start;
661
662 folio_zero_range(folio, zero_start, zero_len);
663 }
664 }
665
666 /* Update page status and unlock. */
667 end_page_read(folio_page(folio, 0), uptodate, start, len);
668 endio_readpage_release_extent(&processed, BTRFS_I(inode),
669 start, end, uptodate);
670
671 ASSERT(bio_offset + len > bio_offset);
672 bio_offset += len;
673
674 }
675 /* Release the last extent */
676 endio_readpage_release_extent(&processed, NULL, 0, 0, false);
677 bio_put(bio);
678}
679
680/*
681 * Populate every free slot in a provided array with pages.
682 *
683 * @nr_pages: number of pages to allocate
684 * @page_array: the array to fill with pages; any existing non-null entries in
685 * the array will be skipped
686 * @extra_gfp: the extra GFP flags for the allocation.
687 *
688 * Return: 0 if all pages were able to be allocated;
689 * -ENOMEM otherwise, the partially allocated pages would be freed and
690 * the array slots zeroed
691 */
692int btrfs_alloc_page_array(unsigned int nr_pages, struct page **page_array,
693 gfp_t extra_gfp)
694{
695 unsigned int allocated;
696
697 for (allocated = 0; allocated < nr_pages;) {
698 unsigned int last = allocated;
699
700 allocated = alloc_pages_bulk_array(GFP_NOFS | extra_gfp,
701 nr_pages, page_array);
702
703 if (allocated == nr_pages)
704 return 0;
705
706 /*
707 * During this iteration, no page could be allocated, even
708 * though alloc_pages_bulk_array() falls back to alloc_page()
709 * if it could not bulk-allocate. So we must be out of memory.
710 */
711 if (allocated == last) {
712 for (int i = 0; i < allocated; i++) {
713 __free_page(page_array[i]);
714 page_array[i] = NULL;
715 }
716 return -ENOMEM;
717 }
718
719 memalloc_retry_wait(GFP_NOFS);
720 }
721 return 0;
722}
723
724/*
725 * Populate needed folios for the extent buffer.
726 *
727 * For now, the folios populated are always in order 0 (aka, single page).
728 */
729static int alloc_eb_folio_array(struct extent_buffer *eb, gfp_t extra_gfp)
730{
731 struct page *page_array[INLINE_EXTENT_BUFFER_PAGES] = { 0 };
732 int num_pages = num_extent_pages(eb);
733 int ret;
734
735 ret = btrfs_alloc_page_array(num_pages, page_array, extra_gfp);
736 if (ret < 0)
737 return ret;
738
739 for (int i = 0; i < num_pages; i++)
740 eb->folios[i] = page_folio(page_array[i]);
741 return 0;
742}
743
744static bool btrfs_bio_is_contig(struct btrfs_bio_ctrl *bio_ctrl,
745 struct page *page, u64 disk_bytenr,
746 unsigned int pg_offset)
747{
748 struct bio *bio = &bio_ctrl->bbio->bio;
749 struct bio_vec *bvec = bio_last_bvec_all(bio);
750 const sector_t sector = disk_bytenr >> SECTOR_SHIFT;
751
752 if (bio_ctrl->compress_type != BTRFS_COMPRESS_NONE) {
753 /*
754 * For compression, all IO should have its logical bytenr set
755 * to the starting bytenr of the compressed extent.
756 */
757 return bio->bi_iter.bi_sector == sector;
758 }
759
760 /*
761 * The contig check requires the following conditions to be met:
762 *
763 * 1) The pages are belonging to the same inode
764 * This is implied by the call chain.
765 *
766 * 2) The range has adjacent logical bytenr
767 *
768 * 3) The range has adjacent file offset
769 * This is required for the usage of btrfs_bio->file_offset.
770 */
771 return bio_end_sector(bio) == sector &&
772 page_offset(bvec->bv_page) + bvec->bv_offset + bvec->bv_len ==
773 page_offset(page) + pg_offset;
774}
775
776static void alloc_new_bio(struct btrfs_inode *inode,
777 struct btrfs_bio_ctrl *bio_ctrl,
778 u64 disk_bytenr, u64 file_offset)
779{
780 struct btrfs_fs_info *fs_info = inode->root->fs_info;
781 struct btrfs_bio *bbio;
782
783 bbio = btrfs_bio_alloc(BIO_MAX_VECS, bio_ctrl->opf, fs_info,
784 bio_ctrl->end_io_func, NULL);
785 bbio->bio.bi_iter.bi_sector = disk_bytenr >> SECTOR_SHIFT;
786 bbio->inode = inode;
787 bbio->file_offset = file_offset;
788 bio_ctrl->bbio = bbio;
789 bio_ctrl->len_to_oe_boundary = U32_MAX;
790
791 /* Limit data write bios to the ordered boundary. */
792 if (bio_ctrl->wbc) {
793 struct btrfs_ordered_extent *ordered;
794
795 ordered = btrfs_lookup_ordered_extent(inode, file_offset);
796 if (ordered) {
797 bio_ctrl->len_to_oe_boundary = min_t(u32, U32_MAX,
798 ordered->file_offset +
799 ordered->disk_num_bytes - file_offset);
800 bbio->ordered = ordered;
801 }
802
803 /*
804 * Pick the last added device to support cgroup writeback. For
805 * multi-device file systems this means blk-cgroup policies have
806 * to always be set on the last added/replaced device.
807 * This is a bit odd but has been like that for a long time.
808 */
809 bio_set_dev(&bbio->bio, fs_info->fs_devices->latest_dev->bdev);
810 wbc_init_bio(bio_ctrl->wbc, &bbio->bio);
811 }
812}
813
814/*
815 * @disk_bytenr: logical bytenr where the write will be
816 * @page: page to add to the bio
817 * @size: portion of page that we want to write to
818 * @pg_offset: offset of the new bio or to check whether we are adding
819 * a contiguous page to the previous one
820 *
821 * The will either add the page into the existing @bio_ctrl->bbio, or allocate a
822 * new one in @bio_ctrl->bbio.
823 * The mirror number for this IO should already be initizlied in
824 * @bio_ctrl->mirror_num.
825 */
826static void submit_extent_page(struct btrfs_bio_ctrl *bio_ctrl,
827 u64 disk_bytenr, struct page *page,
828 size_t size, unsigned long pg_offset)
829{
830 struct btrfs_inode *inode = BTRFS_I(page->mapping->host);
831
832 ASSERT(pg_offset + size <= PAGE_SIZE);
833 ASSERT(bio_ctrl->end_io_func);
834
835 if (bio_ctrl->bbio &&
836 !btrfs_bio_is_contig(bio_ctrl, page, disk_bytenr, pg_offset))
837 submit_one_bio(bio_ctrl);
838
839 do {
840 u32 len = size;
841
842 /* Allocate new bio if needed */
843 if (!bio_ctrl->bbio) {
844 alloc_new_bio(inode, bio_ctrl, disk_bytenr,
845 page_offset(page) + pg_offset);
846 }
847
848 /* Cap to the current ordered extent boundary if there is one. */
849 if (len > bio_ctrl->len_to_oe_boundary) {
850 ASSERT(bio_ctrl->compress_type == BTRFS_COMPRESS_NONE);
851 ASSERT(is_data_inode(&inode->vfs_inode));
852 len = bio_ctrl->len_to_oe_boundary;
853 }
854
855 if (bio_add_page(&bio_ctrl->bbio->bio, page, len, pg_offset) != len) {
856 /* bio full: move on to a new one */
857 submit_one_bio(bio_ctrl);
858 continue;
859 }
860
861 if (bio_ctrl->wbc)
862 wbc_account_cgroup_owner(bio_ctrl->wbc, page, len);
863
864 size -= len;
865 pg_offset += len;
866 disk_bytenr += len;
867
868 /*
869 * len_to_oe_boundary defaults to U32_MAX, which isn't page or
870 * sector aligned. alloc_new_bio() then sets it to the end of
871 * our ordered extent for writes into zoned devices.
872 *
873 * When len_to_oe_boundary is tracking an ordered extent, we
874 * trust the ordered extent code to align things properly, and
875 * the check above to cap our write to the ordered extent
876 * boundary is correct.
877 *
878 * When len_to_oe_boundary is U32_MAX, the cap above would
879 * result in a 4095 byte IO for the last page right before
880 * we hit the bio limit of UINT_MAX. bio_add_page() has all
881 * the checks required to make sure we don't overflow the bio,
882 * and we should just ignore len_to_oe_boundary completely
883 * unless we're using it to track an ordered extent.
884 *
885 * It's pretty hard to make a bio sized U32_MAX, but it can
886 * happen when the page cache is able to feed us contiguous
887 * pages for large extents.
888 */
889 if (bio_ctrl->len_to_oe_boundary != U32_MAX)
890 bio_ctrl->len_to_oe_boundary -= len;
891
892 /* Ordered extent boundary: move on to a new bio. */
893 if (bio_ctrl->len_to_oe_boundary == 0)
894 submit_one_bio(bio_ctrl);
895 } while (size);
896}
897
898static int attach_extent_buffer_folio(struct extent_buffer *eb,
899 struct folio *folio,
900 struct btrfs_subpage *prealloc)
901{
902 struct btrfs_fs_info *fs_info = eb->fs_info;
903 int ret = 0;
904
905 /*
906 * If the page is mapped to btree inode, we should hold the private
907 * lock to prevent race.
908 * For cloned or dummy extent buffers, their pages are not mapped and
909 * will not race with any other ebs.
910 */
911 if (folio->mapping)
912 lockdep_assert_held(&folio->mapping->i_private_lock);
913
914 if (fs_info->nodesize >= PAGE_SIZE) {
915 if (!folio_test_private(folio))
916 folio_attach_private(folio, eb);
917 else
918 WARN_ON(folio_get_private(folio) != eb);
919 return 0;
920 }
921
922 /* Already mapped, just free prealloc */
923 if (folio_test_private(folio)) {
924 btrfs_free_subpage(prealloc);
925 return 0;
926 }
927
928 if (prealloc)
929 /* Has preallocated memory for subpage */
930 folio_attach_private(folio, prealloc);
931 else
932 /* Do new allocation to attach subpage */
933 ret = btrfs_attach_subpage(fs_info, folio, BTRFS_SUBPAGE_METADATA);
934 return ret;
935}
936
937int set_page_extent_mapped(struct page *page)
938{
939 struct folio *folio = page_folio(page);
940 struct btrfs_fs_info *fs_info;
941
942 ASSERT(page->mapping);
943
944 if (folio_test_private(folio))
945 return 0;
946
947 fs_info = btrfs_sb(page->mapping->host->i_sb);
948
949 if (btrfs_is_subpage(fs_info, page->mapping))
950 return btrfs_attach_subpage(fs_info, folio, BTRFS_SUBPAGE_DATA);
951
952 folio_attach_private(folio, (void *)EXTENT_FOLIO_PRIVATE);
953 return 0;
954}
955
956void clear_page_extent_mapped(struct page *page)
957{
958 struct folio *folio = page_folio(page);
959 struct btrfs_fs_info *fs_info;
960
961 ASSERT(page->mapping);
962
963 if (!folio_test_private(folio))
964 return;
965
966 fs_info = btrfs_sb(page->mapping->host->i_sb);
967 if (btrfs_is_subpage(fs_info, page->mapping))
968 return btrfs_detach_subpage(fs_info, folio);
969
970 folio_detach_private(folio);
971}
972
973static struct extent_map *
974__get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
975 u64 start, u64 len, struct extent_map **em_cached)
976{
977 struct extent_map *em;
978
979 if (em_cached && *em_cached) {
980 em = *em_cached;
981 if (extent_map_in_tree(em) && start >= em->start &&
982 start < extent_map_end(em)) {
983 refcount_inc(&em->refs);
984 return em;
985 }
986
987 free_extent_map(em);
988 *em_cached = NULL;
989 }
990
991 em = btrfs_get_extent(BTRFS_I(inode), page, pg_offset, start, len);
992 if (em_cached && !IS_ERR(em)) {
993 BUG_ON(*em_cached);
994 refcount_inc(&em->refs);
995 *em_cached = em;
996 }
997 return em;
998}
999/*
1000 * basic readpage implementation. Locked extent state structs are inserted
1001 * into the tree that are removed when the IO is done (by the end_io
1002 * handlers)
1003 * XXX JDM: This needs looking at to ensure proper page locking
1004 * return 0 on success, otherwise return error
1005 */
1006static int btrfs_do_readpage(struct page *page, struct extent_map **em_cached,
1007 struct btrfs_bio_ctrl *bio_ctrl, u64 *prev_em_start)
1008{
1009 struct inode *inode = page->mapping->host;
1010 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1011 u64 start = page_offset(page);
1012 const u64 end = start + PAGE_SIZE - 1;
1013 u64 cur = start;
1014 u64 extent_offset;
1015 u64 last_byte = i_size_read(inode);
1016 u64 block_start;
1017 struct extent_map *em;
1018 int ret = 0;
1019 size_t pg_offset = 0;
1020 size_t iosize;
1021 size_t blocksize = inode->i_sb->s_blocksize;
1022 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1023
1024 ret = set_page_extent_mapped(page);
1025 if (ret < 0) {
1026 unlock_extent(tree, start, end, NULL);
1027 unlock_page(page);
1028 return ret;
1029 }
1030
1031 if (page->index == last_byte >> PAGE_SHIFT) {
1032 size_t zero_offset = offset_in_page(last_byte);
1033
1034 if (zero_offset) {
1035 iosize = PAGE_SIZE - zero_offset;
1036 memzero_page(page, zero_offset, iosize);
1037 }
1038 }
1039 bio_ctrl->end_io_func = end_bbio_data_read;
1040 begin_page_read(fs_info, page);
1041 while (cur <= end) {
1042 enum btrfs_compression_type compress_type = BTRFS_COMPRESS_NONE;
1043 bool force_bio_submit = false;
1044 u64 disk_bytenr;
1045
1046 ASSERT(IS_ALIGNED(cur, fs_info->sectorsize));
1047 if (cur >= last_byte) {
1048 iosize = PAGE_SIZE - pg_offset;
1049 memzero_page(page, pg_offset, iosize);
1050 unlock_extent(tree, cur, cur + iosize - 1, NULL);
1051 end_page_read(page, true, cur, iosize);
1052 break;
1053 }
1054 em = __get_extent_map(inode, page, pg_offset, cur,
1055 end - cur + 1, em_cached);
1056 if (IS_ERR(em)) {
1057 unlock_extent(tree, cur, end, NULL);
1058 end_page_read(page, false, cur, end + 1 - cur);
1059 return PTR_ERR(em);
1060 }
1061 extent_offset = cur - em->start;
1062 BUG_ON(extent_map_end(em) <= cur);
1063 BUG_ON(end < cur);
1064
1065 compress_type = extent_map_compression(em);
1066
1067 iosize = min(extent_map_end(em) - cur, end - cur + 1);
1068 iosize = ALIGN(iosize, blocksize);
1069 if (compress_type != BTRFS_COMPRESS_NONE)
1070 disk_bytenr = em->block_start;
1071 else
1072 disk_bytenr = em->block_start + extent_offset;
1073 block_start = em->block_start;
1074 if (em->flags & EXTENT_FLAG_PREALLOC)
1075 block_start = EXTENT_MAP_HOLE;
1076
1077 /*
1078 * If we have a file range that points to a compressed extent
1079 * and it's followed by a consecutive file range that points
1080 * to the same compressed extent (possibly with a different
1081 * offset and/or length, so it either points to the whole extent
1082 * or only part of it), we must make sure we do not submit a
1083 * single bio to populate the pages for the 2 ranges because
1084 * this makes the compressed extent read zero out the pages
1085 * belonging to the 2nd range. Imagine the following scenario:
1086 *
1087 * File layout
1088 * [0 - 8K] [8K - 24K]
1089 * | |
1090 * | |
1091 * points to extent X, points to extent X,
1092 * offset 4K, length of 8K offset 0, length 16K
1093 *
1094 * [extent X, compressed length = 4K uncompressed length = 16K]
1095 *
1096 * If the bio to read the compressed extent covers both ranges,
1097 * it will decompress extent X into the pages belonging to the
1098 * first range and then it will stop, zeroing out the remaining
1099 * pages that belong to the other range that points to extent X.
1100 * So here we make sure we submit 2 bios, one for the first
1101 * range and another one for the third range. Both will target
1102 * the same physical extent from disk, but we can't currently
1103 * make the compressed bio endio callback populate the pages
1104 * for both ranges because each compressed bio is tightly
1105 * coupled with a single extent map, and each range can have
1106 * an extent map with a different offset value relative to the
1107 * uncompressed data of our extent and different lengths. This
1108 * is a corner case so we prioritize correctness over
1109 * non-optimal behavior (submitting 2 bios for the same extent).
1110 */
1111 if (compress_type != BTRFS_COMPRESS_NONE &&
1112 prev_em_start && *prev_em_start != (u64)-1 &&
1113 *prev_em_start != em->start)
1114 force_bio_submit = true;
1115
1116 if (prev_em_start)
1117 *prev_em_start = em->start;
1118
1119 free_extent_map(em);
1120 em = NULL;
1121
1122 /* we've found a hole, just zero and go on */
1123 if (block_start == EXTENT_MAP_HOLE) {
1124 memzero_page(page, pg_offset, iosize);
1125
1126 unlock_extent(tree, cur, cur + iosize - 1, NULL);
1127 end_page_read(page, true, cur, iosize);
1128 cur = cur + iosize;
1129 pg_offset += iosize;
1130 continue;
1131 }
1132 /* the get_extent function already copied into the page */
1133 if (block_start == EXTENT_MAP_INLINE) {
1134 unlock_extent(tree, cur, cur + iosize - 1, NULL);
1135 end_page_read(page, true, cur, iosize);
1136 cur = cur + iosize;
1137 pg_offset += iosize;
1138 continue;
1139 }
1140
1141 if (bio_ctrl->compress_type != compress_type) {
1142 submit_one_bio(bio_ctrl);
1143 bio_ctrl->compress_type = compress_type;
1144 }
1145
1146 if (force_bio_submit)
1147 submit_one_bio(bio_ctrl);
1148 submit_extent_page(bio_ctrl, disk_bytenr, page, iosize,
1149 pg_offset);
1150 cur = cur + iosize;
1151 pg_offset += iosize;
1152 }
1153
1154 return 0;
1155}
1156
1157int btrfs_read_folio(struct file *file, struct folio *folio)
1158{
1159 struct page *page = &folio->page;
1160 struct btrfs_inode *inode = BTRFS_I(page->mapping->host);
1161 u64 start = page_offset(page);
1162 u64 end = start + PAGE_SIZE - 1;
1163 struct btrfs_bio_ctrl bio_ctrl = { .opf = REQ_OP_READ };
1164 int ret;
1165
1166 btrfs_lock_and_flush_ordered_range(inode, start, end, NULL);
1167
1168 ret = btrfs_do_readpage(page, NULL, &bio_ctrl, NULL);
1169 /*
1170 * If btrfs_do_readpage() failed we will want to submit the assembled
1171 * bio to do the cleanup.
1172 */
1173 submit_one_bio(&bio_ctrl);
1174 return ret;
1175}
1176
1177static inline void contiguous_readpages(struct page *pages[], int nr_pages,
1178 u64 start, u64 end,
1179 struct extent_map **em_cached,
1180 struct btrfs_bio_ctrl *bio_ctrl,
1181 u64 *prev_em_start)
1182{
1183 struct btrfs_inode *inode = BTRFS_I(pages[0]->mapping->host);
1184 int index;
1185
1186 btrfs_lock_and_flush_ordered_range(inode, start, end, NULL);
1187
1188 for (index = 0; index < nr_pages; index++) {
1189 btrfs_do_readpage(pages[index], em_cached, bio_ctrl,
1190 prev_em_start);
1191 put_page(pages[index]);
1192 }
1193}
1194
1195/*
1196 * helper for __extent_writepage, doing all of the delayed allocation setup.
1197 *
1198 * This returns 1 if btrfs_run_delalloc_range function did all the work required
1199 * to write the page (copy into inline extent). In this case the IO has
1200 * been started and the page is already unlocked.
1201 *
1202 * This returns 0 if all went well (page still locked)
1203 * This returns < 0 if there were errors (page still locked)
1204 */
1205static noinline_for_stack int writepage_delalloc(struct btrfs_inode *inode,
1206 struct page *page, struct writeback_control *wbc)
1207{
1208 const u64 page_start = page_offset(page);
1209 const u64 page_end = page_start + PAGE_SIZE - 1;
1210 u64 delalloc_start = page_start;
1211 u64 delalloc_end = page_end;
1212 u64 delalloc_to_write = 0;
1213 int ret = 0;
1214
1215 while (delalloc_start < page_end) {
1216 delalloc_end = page_end;
1217 if (!find_lock_delalloc_range(&inode->vfs_inode, page,
1218 &delalloc_start, &delalloc_end)) {
1219 delalloc_start = delalloc_end + 1;
1220 continue;
1221 }
1222
1223 ret = btrfs_run_delalloc_range(inode, page, delalloc_start,
1224 delalloc_end, wbc);
1225 if (ret < 0)
1226 return ret;
1227
1228 delalloc_start = delalloc_end + 1;
1229 }
1230
1231 /*
1232 * delalloc_end is already one less than the total length, so
1233 * we don't subtract one from PAGE_SIZE
1234 */
1235 delalloc_to_write +=
1236 DIV_ROUND_UP(delalloc_end + 1 - page_start, PAGE_SIZE);
1237
1238 /*
1239 * If btrfs_run_dealloc_range() already started I/O and unlocked
1240 * the pages, we just need to account for them here.
1241 */
1242 if (ret == 1) {
1243 wbc->nr_to_write -= delalloc_to_write;
1244 return 1;
1245 }
1246
1247 if (wbc->nr_to_write < delalloc_to_write) {
1248 int thresh = 8192;
1249
1250 if (delalloc_to_write < thresh * 2)
1251 thresh = delalloc_to_write;
1252 wbc->nr_to_write = min_t(u64, delalloc_to_write,
1253 thresh);
1254 }
1255
1256 return 0;
1257}
1258
1259/*
1260 * Find the first byte we need to write.
1261 *
1262 * For subpage, one page can contain several sectors, and
1263 * __extent_writepage_io() will just grab all extent maps in the page
1264 * range and try to submit all non-inline/non-compressed extents.
1265 *
1266 * This is a big problem for subpage, we shouldn't re-submit already written
1267 * data at all.
1268 * This function will lookup subpage dirty bit to find which range we really
1269 * need to submit.
1270 *
1271 * Return the next dirty range in [@start, @end).
1272 * If no dirty range is found, @start will be page_offset(page) + PAGE_SIZE.
1273 */
1274static void find_next_dirty_byte(struct btrfs_fs_info *fs_info,
1275 struct page *page, u64 *start, u64 *end)
1276{
1277 struct folio *folio = page_folio(page);
1278 struct btrfs_subpage *subpage = folio_get_private(folio);
1279 struct btrfs_subpage_info *spi = fs_info->subpage_info;
1280 u64 orig_start = *start;
1281 /* Declare as unsigned long so we can use bitmap ops */
1282 unsigned long flags;
1283 int range_start_bit;
1284 int range_end_bit;
1285
1286 /*
1287 * For regular sector size == page size case, since one page only
1288 * contains one sector, we return the page offset directly.
1289 */
1290 if (!btrfs_is_subpage(fs_info, page->mapping)) {
1291 *start = page_offset(page);
1292 *end = page_offset(page) + PAGE_SIZE;
1293 return;
1294 }
1295
1296 range_start_bit = spi->dirty_offset +
1297 (offset_in_page(orig_start) >> fs_info->sectorsize_bits);
1298
1299 /* We should have the page locked, but just in case */
1300 spin_lock_irqsave(&subpage->lock, flags);
1301 bitmap_next_set_region(subpage->bitmaps, &range_start_bit, &range_end_bit,
1302 spi->dirty_offset + spi->bitmap_nr_bits);
1303 spin_unlock_irqrestore(&subpage->lock, flags);
1304
1305 range_start_bit -= spi->dirty_offset;
1306 range_end_bit -= spi->dirty_offset;
1307
1308 *start = page_offset(page) + range_start_bit * fs_info->sectorsize;
1309 *end = page_offset(page) + range_end_bit * fs_info->sectorsize;
1310}
1311
1312/*
1313 * helper for __extent_writepage. This calls the writepage start hooks,
1314 * and does the loop to map the page into extents and bios.
1315 *
1316 * We return 1 if the IO is started and the page is unlocked,
1317 * 0 if all went well (page still locked)
1318 * < 0 if there were errors (page still locked)
1319 */
1320static noinline_for_stack int __extent_writepage_io(struct btrfs_inode *inode,
1321 struct page *page,
1322 struct btrfs_bio_ctrl *bio_ctrl,
1323 loff_t i_size,
1324 int *nr_ret)
1325{
1326 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1327 u64 cur = page_offset(page);
1328 u64 end = cur + PAGE_SIZE - 1;
1329 u64 extent_offset;
1330 u64 block_start;
1331 struct extent_map *em;
1332 int ret = 0;
1333 int nr = 0;
1334
1335 ret = btrfs_writepage_cow_fixup(page);
1336 if (ret) {
1337 /* Fixup worker will requeue */
1338 redirty_page_for_writepage(bio_ctrl->wbc, page);
1339 unlock_page(page);
1340 return 1;
1341 }
1342
1343 bio_ctrl->end_io_func = end_bbio_data_write;
1344 while (cur <= end) {
1345 u32 len = end - cur + 1;
1346 u64 disk_bytenr;
1347 u64 em_end;
1348 u64 dirty_range_start = cur;
1349 u64 dirty_range_end;
1350 u32 iosize;
1351
1352 if (cur >= i_size) {
1353 btrfs_mark_ordered_io_finished(inode, page, cur, len,
1354 true);
1355 /*
1356 * This range is beyond i_size, thus we don't need to
1357 * bother writing back.
1358 * But we still need to clear the dirty subpage bit, or
1359 * the next time the page gets dirtied, we will try to
1360 * writeback the sectors with subpage dirty bits,
1361 * causing writeback without ordered extent.
1362 */
1363 btrfs_folio_clear_dirty(fs_info, page_folio(page), cur, len);
1364 break;
1365 }
1366
1367 find_next_dirty_byte(fs_info, page, &dirty_range_start,
1368 &dirty_range_end);
1369 if (cur < dirty_range_start) {
1370 cur = dirty_range_start;
1371 continue;
1372 }
1373
1374 em = btrfs_get_extent(inode, NULL, 0, cur, len);
1375 if (IS_ERR(em)) {
1376 ret = PTR_ERR_OR_ZERO(em);
1377 goto out_error;
1378 }
1379
1380 extent_offset = cur - em->start;
1381 em_end = extent_map_end(em);
1382 ASSERT(cur <= em_end);
1383 ASSERT(cur < end);
1384 ASSERT(IS_ALIGNED(em->start, fs_info->sectorsize));
1385 ASSERT(IS_ALIGNED(em->len, fs_info->sectorsize));
1386
1387 block_start = em->block_start;
1388 disk_bytenr = em->block_start + extent_offset;
1389
1390 ASSERT(!extent_map_is_compressed(em));
1391 ASSERT(block_start != EXTENT_MAP_HOLE);
1392 ASSERT(block_start != EXTENT_MAP_INLINE);
1393
1394 /*
1395 * Note that em_end from extent_map_end() and dirty_range_end from
1396 * find_next_dirty_byte() are all exclusive
1397 */
1398 iosize = min(min(em_end, end + 1), dirty_range_end) - cur;
1399 free_extent_map(em);
1400 em = NULL;
1401
1402 btrfs_set_range_writeback(inode, cur, cur + iosize - 1);
1403 if (!PageWriteback(page)) {
1404 btrfs_err(inode->root->fs_info,
1405 "page %lu not writeback, cur %llu end %llu",
1406 page->index, cur, end);
1407 }
1408
1409 /*
1410 * Although the PageDirty bit is cleared before entering this
1411 * function, subpage dirty bit is not cleared.
1412 * So clear subpage dirty bit here so next time we won't submit
1413 * page for range already written to disk.
1414 */
1415 btrfs_folio_clear_dirty(fs_info, page_folio(page), cur, iosize);
1416
1417 submit_extent_page(bio_ctrl, disk_bytenr, page, iosize,
1418 cur - page_offset(page));
1419 cur += iosize;
1420 nr++;
1421 }
1422
1423 btrfs_folio_assert_not_dirty(fs_info, page_folio(page));
1424 *nr_ret = nr;
1425 return 0;
1426
1427out_error:
1428 /*
1429 * If we finish without problem, we should not only clear page dirty,
1430 * but also empty subpage dirty bits
1431 */
1432 *nr_ret = nr;
1433 return ret;
1434}
1435
1436/*
1437 * the writepage semantics are similar to regular writepage. extent
1438 * records are inserted to lock ranges in the tree, and as dirty areas
1439 * are found, they are marked writeback. Then the lock bits are removed
1440 * and the end_io handler clears the writeback ranges
1441 *
1442 * Return 0 if everything goes well.
1443 * Return <0 for error.
1444 */
1445static int __extent_writepage(struct page *page, struct btrfs_bio_ctrl *bio_ctrl)
1446{
1447 struct folio *folio = page_folio(page);
1448 struct inode *inode = page->mapping->host;
1449 const u64 page_start = page_offset(page);
1450 int ret;
1451 int nr = 0;
1452 size_t pg_offset;
1453 loff_t i_size = i_size_read(inode);
1454 unsigned long end_index = i_size >> PAGE_SHIFT;
1455
1456 trace___extent_writepage(page, inode, bio_ctrl->wbc);
1457
1458 WARN_ON(!PageLocked(page));
1459
1460 pg_offset = offset_in_page(i_size);
1461 if (page->index > end_index ||
1462 (page->index == end_index && !pg_offset)) {
1463 folio_invalidate(folio, 0, folio_size(folio));
1464 folio_unlock(folio);
1465 return 0;
1466 }
1467
1468 if (page->index == end_index)
1469 memzero_page(page, pg_offset, PAGE_SIZE - pg_offset);
1470
1471 ret = set_page_extent_mapped(page);
1472 if (ret < 0)
1473 goto done;
1474
1475 ret = writepage_delalloc(BTRFS_I(inode), page, bio_ctrl->wbc);
1476 if (ret == 1)
1477 return 0;
1478 if (ret)
1479 goto done;
1480
1481 ret = __extent_writepage_io(BTRFS_I(inode), page, bio_ctrl, i_size, &nr);
1482 if (ret == 1)
1483 return 0;
1484
1485 bio_ctrl->wbc->nr_to_write--;
1486
1487done:
1488 if (nr == 0) {
1489 /* make sure the mapping tag for page dirty gets cleared */
1490 set_page_writeback(page);
1491 end_page_writeback(page);
1492 }
1493 if (ret) {
1494 btrfs_mark_ordered_io_finished(BTRFS_I(inode), page, page_start,
1495 PAGE_SIZE, !ret);
1496 mapping_set_error(page->mapping, ret);
1497 }
1498 unlock_page(page);
1499 ASSERT(ret <= 0);
1500 return ret;
1501}
1502
1503void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
1504{
1505 wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
1506 TASK_UNINTERRUPTIBLE);
1507}
1508
1509/*
1510 * Lock extent buffer status and pages for writeback.
1511 *
1512 * Return %false if the extent buffer doesn't need to be submitted (e.g. the
1513 * extent buffer is not dirty)
1514 * Return %true is the extent buffer is submitted to bio.
1515 */
1516static noinline_for_stack bool lock_extent_buffer_for_io(struct extent_buffer *eb,
1517 struct writeback_control *wbc)
1518{
1519 struct btrfs_fs_info *fs_info = eb->fs_info;
1520 bool ret = false;
1521
1522 btrfs_tree_lock(eb);
1523 while (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
1524 btrfs_tree_unlock(eb);
1525 if (wbc->sync_mode != WB_SYNC_ALL)
1526 return false;
1527 wait_on_extent_buffer_writeback(eb);
1528 btrfs_tree_lock(eb);
1529 }
1530
1531 /*
1532 * We need to do this to prevent races in people who check if the eb is
1533 * under IO since we can end up having no IO bits set for a short period
1534 * of time.
1535 */
1536 spin_lock(&eb->refs_lock);
1537 if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
1538 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
1539 spin_unlock(&eb->refs_lock);
1540 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
1541 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
1542 -eb->len,
1543 fs_info->dirty_metadata_batch);
1544 ret = true;
1545 } else {
1546 spin_unlock(&eb->refs_lock);
1547 }
1548 btrfs_tree_unlock(eb);
1549 return ret;
1550}
1551
1552static void set_btree_ioerr(struct extent_buffer *eb)
1553{
1554 struct btrfs_fs_info *fs_info = eb->fs_info;
1555
1556 set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
1557
1558 /*
1559 * A read may stumble upon this buffer later, make sure that it gets an
1560 * error and knows there was an error.
1561 */
1562 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
1563
1564 /*
1565 * We need to set the mapping with the io error as well because a write
1566 * error will flip the file system readonly, and then syncfs() will
1567 * return a 0 because we are readonly if we don't modify the err seq for
1568 * the superblock.
1569 */
1570 mapping_set_error(eb->fs_info->btree_inode->i_mapping, -EIO);
1571
1572 /*
1573 * If writeback for a btree extent that doesn't belong to a log tree
1574 * failed, increment the counter transaction->eb_write_errors.
1575 * We do this because while the transaction is running and before it's
1576 * committing (when we call filemap_fdata[write|wait]_range against
1577 * the btree inode), we might have
1578 * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
1579 * returns an error or an error happens during writeback, when we're
1580 * committing the transaction we wouldn't know about it, since the pages
1581 * can be no longer dirty nor marked anymore for writeback (if a
1582 * subsequent modification to the extent buffer didn't happen before the
1583 * transaction commit), which makes filemap_fdata[write|wait]_range not
1584 * able to find the pages tagged with SetPageError at transaction
1585 * commit time. So if this happens we must abort the transaction,
1586 * otherwise we commit a super block with btree roots that point to
1587 * btree nodes/leafs whose content on disk is invalid - either garbage
1588 * or the content of some node/leaf from a past generation that got
1589 * cowed or deleted and is no longer valid.
1590 *
1591 * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
1592 * not be enough - we need to distinguish between log tree extents vs
1593 * non-log tree extents, and the next filemap_fdatawait_range() call
1594 * will catch and clear such errors in the mapping - and that call might
1595 * be from a log sync and not from a transaction commit. Also, checking
1596 * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
1597 * not done and would not be reliable - the eb might have been released
1598 * from memory and reading it back again means that flag would not be
1599 * set (since it's a runtime flag, not persisted on disk).
1600 *
1601 * Using the flags below in the btree inode also makes us achieve the
1602 * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
1603 * writeback for all dirty pages and before filemap_fdatawait_range()
1604 * is called, the writeback for all dirty pages had already finished
1605 * with errors - because we were not using AS_EIO/AS_ENOSPC,
1606 * filemap_fdatawait_range() would return success, as it could not know
1607 * that writeback errors happened (the pages were no longer tagged for
1608 * writeback).
1609 */
1610 switch (eb->log_index) {
1611 case -1:
1612 set_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags);
1613 break;
1614 case 0:
1615 set_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
1616 break;
1617 case 1:
1618 set_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
1619 break;
1620 default:
1621 BUG(); /* unexpected, logic error */
1622 }
1623}
1624
1625/*
1626 * The endio specific version which won't touch any unsafe spinlock in endio
1627 * context.
1628 */
1629static struct extent_buffer *find_extent_buffer_nolock(
1630 struct btrfs_fs_info *fs_info, u64 start)
1631{
1632 struct extent_buffer *eb;
1633
1634 rcu_read_lock();
1635 eb = radix_tree_lookup(&fs_info->buffer_radix,
1636 start >> fs_info->sectorsize_bits);
1637 if (eb && atomic_inc_not_zero(&eb->refs)) {
1638 rcu_read_unlock();
1639 return eb;
1640 }
1641 rcu_read_unlock();
1642 return NULL;
1643}
1644
1645static void end_bbio_meta_write(struct btrfs_bio *bbio)
1646{
1647 struct extent_buffer *eb = bbio->private;
1648 struct btrfs_fs_info *fs_info = eb->fs_info;
1649 bool uptodate = !bbio->bio.bi_status;
1650 struct folio_iter fi;
1651 u32 bio_offset = 0;
1652
1653 if (!uptodate)
1654 set_btree_ioerr(eb);
1655
1656 bio_for_each_folio_all(fi, &bbio->bio) {
1657 u64 start = eb->start + bio_offset;
1658 struct folio *folio = fi.folio;
1659 u32 len = fi.length;
1660
1661 btrfs_folio_clear_writeback(fs_info, folio, start, len);
1662 bio_offset += len;
1663 }
1664
1665 clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
1666 smp_mb__after_atomic();
1667 wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
1668
1669 bio_put(&bbio->bio);
1670}
1671
1672static void prepare_eb_write(struct extent_buffer *eb)
1673{
1674 u32 nritems;
1675 unsigned long start;
1676 unsigned long end;
1677
1678 clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
1679
1680 /* Set btree blocks beyond nritems with 0 to avoid stale content */
1681 nritems = btrfs_header_nritems(eb);
1682 if (btrfs_header_level(eb) > 0) {
1683 end = btrfs_node_key_ptr_offset(eb, nritems);
1684 memzero_extent_buffer(eb, end, eb->len - end);
1685 } else {
1686 /*
1687 * Leaf:
1688 * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
1689 */
1690 start = btrfs_item_nr_offset(eb, nritems);
1691 end = btrfs_item_nr_offset(eb, 0);
1692 if (nritems == 0)
1693 end += BTRFS_LEAF_DATA_SIZE(eb->fs_info);
1694 else
1695 end += btrfs_item_offset(eb, nritems - 1);
1696 memzero_extent_buffer(eb, start, end - start);
1697 }
1698}
1699
1700static noinline_for_stack void write_one_eb(struct extent_buffer *eb,
1701 struct writeback_control *wbc)
1702{
1703 struct btrfs_fs_info *fs_info = eb->fs_info;
1704 struct btrfs_bio *bbio;
1705
1706 prepare_eb_write(eb);
1707
1708 bbio = btrfs_bio_alloc(INLINE_EXTENT_BUFFER_PAGES,
1709 REQ_OP_WRITE | REQ_META | wbc_to_write_flags(wbc),
1710 eb->fs_info, end_bbio_meta_write, eb);
1711 bbio->bio.bi_iter.bi_sector = eb->start >> SECTOR_SHIFT;
1712 bio_set_dev(&bbio->bio, fs_info->fs_devices->latest_dev->bdev);
1713 wbc_init_bio(wbc, &bbio->bio);
1714 bbio->inode = BTRFS_I(eb->fs_info->btree_inode);
1715 bbio->file_offset = eb->start;
1716 if (fs_info->nodesize < PAGE_SIZE) {
1717 struct folio *folio = eb->folios[0];
1718 bool ret;
1719
1720 folio_lock(folio);
1721 btrfs_subpage_set_writeback(fs_info, folio, eb->start, eb->len);
1722 if (btrfs_subpage_clear_and_test_dirty(fs_info, folio, eb->start,
1723 eb->len)) {
1724 folio_clear_dirty_for_io(folio);
1725 wbc->nr_to_write--;
1726 }
1727 ret = bio_add_folio(&bbio->bio, folio, eb->len,
1728 eb->start - folio_pos(folio));
1729 ASSERT(ret);
1730 wbc_account_cgroup_owner(wbc, folio_page(folio, 0), eb->len);
1731 folio_unlock(folio);
1732 } else {
1733 int num_folios = num_extent_folios(eb);
1734
1735 for (int i = 0; i < num_folios; i++) {
1736 struct folio *folio = eb->folios[i];
1737 bool ret;
1738
1739 folio_lock(folio);
1740 folio_clear_dirty_for_io(folio);
1741 folio_start_writeback(folio);
1742 ret = bio_add_folio(&bbio->bio, folio, folio_size(folio), 0);
1743 ASSERT(ret);
1744 wbc_account_cgroup_owner(wbc, folio_page(folio, 0),
1745 folio_size(folio));
1746 wbc->nr_to_write -= folio_nr_pages(folio);
1747 folio_unlock(folio);
1748 }
1749 }
1750 btrfs_submit_bio(bbio, 0);
1751}
1752
1753/*
1754 * Submit one subpage btree page.
1755 *
1756 * The main difference to submit_eb_page() is:
1757 * - Page locking
1758 * For subpage, we don't rely on page locking at all.
1759 *
1760 * - Flush write bio
1761 * We only flush bio if we may be unable to fit current extent buffers into
1762 * current bio.
1763 *
1764 * Return >=0 for the number of submitted extent buffers.
1765 * Return <0 for fatal error.
1766 */
1767static int submit_eb_subpage(struct page *page, struct writeback_control *wbc)
1768{
1769 struct btrfs_fs_info *fs_info = btrfs_sb(page->mapping->host->i_sb);
1770 struct folio *folio = page_folio(page);
1771 int submitted = 0;
1772 u64 page_start = page_offset(page);
1773 int bit_start = 0;
1774 int sectors_per_node = fs_info->nodesize >> fs_info->sectorsize_bits;
1775
1776 /* Lock and write each dirty extent buffers in the range */
1777 while (bit_start < fs_info->subpage_info->bitmap_nr_bits) {
1778 struct btrfs_subpage *subpage = folio_get_private(folio);
1779 struct extent_buffer *eb;
1780 unsigned long flags;
1781 u64 start;
1782
1783 /*
1784 * Take private lock to ensure the subpage won't be detached
1785 * in the meantime.
1786 */
1787 spin_lock(&page->mapping->i_private_lock);
1788 if (!folio_test_private(folio)) {
1789 spin_unlock(&page->mapping->i_private_lock);
1790 break;
1791 }
1792 spin_lock_irqsave(&subpage->lock, flags);
1793 if (!test_bit(bit_start + fs_info->subpage_info->dirty_offset,
1794 subpage->bitmaps)) {
1795 spin_unlock_irqrestore(&subpage->lock, flags);
1796 spin_unlock(&page->mapping->i_private_lock);
1797 bit_start++;
1798 continue;
1799 }
1800
1801 start = page_start + bit_start * fs_info->sectorsize;
1802 bit_start += sectors_per_node;
1803
1804 /*
1805 * Here we just want to grab the eb without touching extra
1806 * spin locks, so call find_extent_buffer_nolock().
1807 */
1808 eb = find_extent_buffer_nolock(fs_info, start);
1809 spin_unlock_irqrestore(&subpage->lock, flags);
1810 spin_unlock(&page->mapping->i_private_lock);
1811
1812 /*
1813 * The eb has already reached 0 refs thus find_extent_buffer()
1814 * doesn't return it. We don't need to write back such eb
1815 * anyway.
1816 */
1817 if (!eb)
1818 continue;
1819
1820 if (lock_extent_buffer_for_io(eb, wbc)) {
1821 write_one_eb(eb, wbc);
1822 submitted++;
1823 }
1824 free_extent_buffer(eb);
1825 }
1826 return submitted;
1827}
1828
1829/*
1830 * Submit all page(s) of one extent buffer.
1831 *
1832 * @page: the page of one extent buffer
1833 * @eb_context: to determine if we need to submit this page, if current page
1834 * belongs to this eb, we don't need to submit
1835 *
1836 * The caller should pass each page in their bytenr order, and here we use
1837 * @eb_context to determine if we have submitted pages of one extent buffer.
1838 *
1839 * If we have, we just skip until we hit a new page that doesn't belong to
1840 * current @eb_context.
1841 *
1842 * If not, we submit all the page(s) of the extent buffer.
1843 *
1844 * Return >0 if we have submitted the extent buffer successfully.
1845 * Return 0 if we don't need to submit the page, as it's already submitted by
1846 * previous call.
1847 * Return <0 for fatal error.
1848 */
1849static int submit_eb_page(struct page *page, struct btrfs_eb_write_context *ctx)
1850{
1851 struct writeback_control *wbc = ctx->wbc;
1852 struct address_space *mapping = page->mapping;
1853 struct folio *folio = page_folio(page);
1854 struct extent_buffer *eb;
1855 int ret;
1856
1857 if (!folio_test_private(folio))
1858 return 0;
1859
1860 if (btrfs_sb(page->mapping->host->i_sb)->nodesize < PAGE_SIZE)
1861 return submit_eb_subpage(page, wbc);
1862
1863 spin_lock(&mapping->i_private_lock);
1864 if (!folio_test_private(folio)) {
1865 spin_unlock(&mapping->i_private_lock);
1866 return 0;
1867 }
1868
1869 eb = folio_get_private(folio);
1870
1871 /*
1872 * Shouldn't happen and normally this would be a BUG_ON but no point
1873 * crashing the machine for something we can survive anyway.
1874 */
1875 if (WARN_ON(!eb)) {
1876 spin_unlock(&mapping->i_private_lock);
1877 return 0;
1878 }
1879
1880 if (eb == ctx->eb) {
1881 spin_unlock(&mapping->i_private_lock);
1882 return 0;
1883 }
1884 ret = atomic_inc_not_zero(&eb->refs);
1885 spin_unlock(&mapping->i_private_lock);
1886 if (!ret)
1887 return 0;
1888
1889 ctx->eb = eb;
1890
1891 ret = btrfs_check_meta_write_pointer(eb->fs_info, ctx);
1892 if (ret) {
1893 if (ret == -EBUSY)
1894 ret = 0;
1895 free_extent_buffer(eb);
1896 return ret;
1897 }
1898
1899 if (!lock_extent_buffer_for_io(eb, wbc)) {
1900 free_extent_buffer(eb);
1901 return 0;
1902 }
1903 /* Implies write in zoned mode. */
1904 if (ctx->zoned_bg) {
1905 /* Mark the last eb in the block group. */
1906 btrfs_schedule_zone_finish_bg(ctx->zoned_bg, eb);
1907 ctx->zoned_bg->meta_write_pointer += eb->len;
1908 }
1909 write_one_eb(eb, wbc);
1910 free_extent_buffer(eb);
1911 return 1;
1912}
1913
1914int btree_write_cache_pages(struct address_space *mapping,
1915 struct writeback_control *wbc)
1916{
1917 struct btrfs_eb_write_context ctx = { .wbc = wbc };
1918 struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
1919 int ret = 0;
1920 int done = 0;
1921 int nr_to_write_done = 0;
1922 struct folio_batch fbatch;
1923 unsigned int nr_folios;
1924 pgoff_t index;
1925 pgoff_t end; /* Inclusive */
1926 int scanned = 0;
1927 xa_mark_t tag;
1928
1929 folio_batch_init(&fbatch);
1930 if (wbc->range_cyclic) {
1931 index = mapping->writeback_index; /* Start from prev offset */
1932 end = -1;
1933 /*
1934 * Start from the beginning does not need to cycle over the
1935 * range, mark it as scanned.
1936 */
1937 scanned = (index == 0);
1938 } else {
1939 index = wbc->range_start >> PAGE_SHIFT;
1940 end = wbc->range_end >> PAGE_SHIFT;
1941 scanned = 1;
1942 }
1943 if (wbc->sync_mode == WB_SYNC_ALL)
1944 tag = PAGECACHE_TAG_TOWRITE;
1945 else
1946 tag = PAGECACHE_TAG_DIRTY;
1947 btrfs_zoned_meta_io_lock(fs_info);
1948retry:
1949 if (wbc->sync_mode == WB_SYNC_ALL)
1950 tag_pages_for_writeback(mapping, index, end);
1951 while (!done && !nr_to_write_done && (index <= end) &&
1952 (nr_folios = filemap_get_folios_tag(mapping, &index, end,
1953 tag, &fbatch))) {
1954 unsigned i;
1955
1956 for (i = 0; i < nr_folios; i++) {
1957 struct folio *folio = fbatch.folios[i];
1958
1959 ret = submit_eb_page(&folio->page, &ctx);
1960 if (ret == 0)
1961 continue;
1962 if (ret < 0) {
1963 done = 1;
1964 break;
1965 }
1966
1967 /*
1968 * the filesystem may choose to bump up nr_to_write.
1969 * We have to make sure to honor the new nr_to_write
1970 * at any time
1971 */
1972 nr_to_write_done = wbc->nr_to_write <= 0;
1973 }
1974 folio_batch_release(&fbatch);
1975 cond_resched();
1976 }
1977 if (!scanned && !done) {
1978 /*
1979 * We hit the last page and there is more work to be done: wrap
1980 * back to the start of the file
1981 */
1982 scanned = 1;
1983 index = 0;
1984 goto retry;
1985 }
1986 /*
1987 * If something went wrong, don't allow any metadata write bio to be
1988 * submitted.
1989 *
1990 * This would prevent use-after-free if we had dirty pages not
1991 * cleaned up, which can still happen by fuzzed images.
1992 *
1993 * - Bad extent tree
1994 * Allowing existing tree block to be allocated for other trees.
1995 *
1996 * - Log tree operations
1997 * Exiting tree blocks get allocated to log tree, bumps its
1998 * generation, then get cleaned in tree re-balance.
1999 * Such tree block will not be written back, since it's clean,
2000 * thus no WRITTEN flag set.
2001 * And after log writes back, this tree block is not traced by
2002 * any dirty extent_io_tree.
2003 *
2004 * - Offending tree block gets re-dirtied from its original owner
2005 * Since it has bumped generation, no WRITTEN flag, it can be
2006 * reused without COWing. This tree block will not be traced
2007 * by btrfs_transaction::dirty_pages.
2008 *
2009 * Now such dirty tree block will not be cleaned by any dirty
2010 * extent io tree. Thus we don't want to submit such wild eb
2011 * if the fs already has error.
2012 *
2013 * We can get ret > 0 from submit_extent_page() indicating how many ebs
2014 * were submitted. Reset it to 0 to avoid false alerts for the caller.
2015 */
2016 if (ret > 0)
2017 ret = 0;
2018 if (!ret && BTRFS_FS_ERROR(fs_info))
2019 ret = -EROFS;
2020
2021 if (ctx.zoned_bg)
2022 btrfs_put_block_group(ctx.zoned_bg);
2023 btrfs_zoned_meta_io_unlock(fs_info);
2024 return ret;
2025}
2026
2027/*
2028 * Walk the list of dirty pages of the given address space and write all of them.
2029 *
2030 * @mapping: address space structure to write
2031 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2032 * @bio_ctrl: holds context for the write, namely the bio
2033 *
2034 * If a page is already under I/O, write_cache_pages() skips it, even
2035 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2036 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2037 * and msync() need to guarantee that all the data which was dirty at the time
2038 * the call was made get new I/O started against them. If wbc->sync_mode is
2039 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2040 * existing IO to complete.
2041 */
2042static int extent_write_cache_pages(struct address_space *mapping,
2043 struct btrfs_bio_ctrl *bio_ctrl)
2044{
2045 struct writeback_control *wbc = bio_ctrl->wbc;
2046 struct inode *inode = mapping->host;
2047 int ret = 0;
2048 int done = 0;
2049 int nr_to_write_done = 0;
2050 struct folio_batch fbatch;
2051 unsigned int nr_folios;
2052 pgoff_t index;
2053 pgoff_t end; /* Inclusive */
2054 pgoff_t done_index;
2055 int range_whole = 0;
2056 int scanned = 0;
2057 xa_mark_t tag;
2058
2059 /*
2060 * We have to hold onto the inode so that ordered extents can do their
2061 * work when the IO finishes. The alternative to this is failing to add
2062 * an ordered extent if the igrab() fails there and that is a huge pain
2063 * to deal with, so instead just hold onto the inode throughout the
2064 * writepages operation. If it fails here we are freeing up the inode
2065 * anyway and we'd rather not waste our time writing out stuff that is
2066 * going to be truncated anyway.
2067 */
2068 if (!igrab(inode))
2069 return 0;
2070
2071 folio_batch_init(&fbatch);
2072 if (wbc->range_cyclic) {
2073 index = mapping->writeback_index; /* Start from prev offset */
2074 end = -1;
2075 /*
2076 * Start from the beginning does not need to cycle over the
2077 * range, mark it as scanned.
2078 */
2079 scanned = (index == 0);
2080 } else {
2081 index = wbc->range_start >> PAGE_SHIFT;
2082 end = wbc->range_end >> PAGE_SHIFT;
2083 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2084 range_whole = 1;
2085 scanned = 1;
2086 }
2087
2088 /*
2089 * We do the tagged writepage as long as the snapshot flush bit is set
2090 * and we are the first one who do the filemap_flush() on this inode.
2091 *
2092 * The nr_to_write == LONG_MAX is needed to make sure other flushers do
2093 * not race in and drop the bit.
2094 */
2095 if (range_whole && wbc->nr_to_write == LONG_MAX &&
2096 test_and_clear_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
2097 &BTRFS_I(inode)->runtime_flags))
2098 wbc->tagged_writepages = 1;
2099
2100 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2101 tag = PAGECACHE_TAG_TOWRITE;
2102 else
2103 tag = PAGECACHE_TAG_DIRTY;
2104retry:
2105 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2106 tag_pages_for_writeback(mapping, index, end);
2107 done_index = index;
2108 while (!done && !nr_to_write_done && (index <= end) &&
2109 (nr_folios = filemap_get_folios_tag(mapping, &index,
2110 end, tag, &fbatch))) {
2111 unsigned i;
2112
2113 for (i = 0; i < nr_folios; i++) {
2114 struct folio *folio = fbatch.folios[i];
2115
2116 done_index = folio_next_index(folio);
2117 /*
2118 * At this point we hold neither the i_pages lock nor
2119 * the page lock: the page may be truncated or
2120 * invalidated (changing page->mapping to NULL),
2121 * or even swizzled back from swapper_space to
2122 * tmpfs file mapping
2123 */
2124 if (!folio_trylock(folio)) {
2125 submit_write_bio(bio_ctrl, 0);
2126 folio_lock(folio);
2127 }
2128
2129 if (unlikely(folio->mapping != mapping)) {
2130 folio_unlock(folio);
2131 continue;
2132 }
2133
2134 if (!folio_test_dirty(folio)) {
2135 /* Someone wrote it for us. */
2136 folio_unlock(folio);
2137 continue;
2138 }
2139
2140 if (wbc->sync_mode != WB_SYNC_NONE) {
2141 if (folio_test_writeback(folio))
2142 submit_write_bio(bio_ctrl, 0);
2143 folio_wait_writeback(folio);
2144 }
2145
2146 if (folio_test_writeback(folio) ||
2147 !folio_clear_dirty_for_io(folio)) {
2148 folio_unlock(folio);
2149 continue;
2150 }
2151
2152 ret = __extent_writepage(&folio->page, bio_ctrl);
2153 if (ret < 0) {
2154 done = 1;
2155 break;
2156 }
2157
2158 /*
2159 * The filesystem may choose to bump up nr_to_write.
2160 * We have to make sure to honor the new nr_to_write
2161 * at any time.
2162 */
2163 nr_to_write_done = (wbc->sync_mode == WB_SYNC_NONE &&
2164 wbc->nr_to_write <= 0);
2165 }
2166 folio_batch_release(&fbatch);
2167 cond_resched();
2168 }
2169 if (!scanned && !done) {
2170 /*
2171 * We hit the last page and there is more work to be done: wrap
2172 * back to the start of the file
2173 */
2174 scanned = 1;
2175 index = 0;
2176
2177 /*
2178 * If we're looping we could run into a page that is locked by a
2179 * writer and that writer could be waiting on writeback for a
2180 * page in our current bio, and thus deadlock, so flush the
2181 * write bio here.
2182 */
2183 submit_write_bio(bio_ctrl, 0);
2184 goto retry;
2185 }
2186
2187 if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
2188 mapping->writeback_index = done_index;
2189
2190 btrfs_add_delayed_iput(BTRFS_I(inode));
2191 return ret;
2192}
2193
2194/*
2195 * Submit the pages in the range to bio for call sites which delalloc range has
2196 * already been ran (aka, ordered extent inserted) and all pages are still
2197 * locked.
2198 */
2199void extent_write_locked_range(struct inode *inode, struct page *locked_page,
2200 u64 start, u64 end, struct writeback_control *wbc,
2201 bool pages_dirty)
2202{
2203 bool found_error = false;
2204 int ret = 0;
2205 struct address_space *mapping = inode->i_mapping;
2206 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2207 const u32 sectorsize = fs_info->sectorsize;
2208 loff_t i_size = i_size_read(inode);
2209 u64 cur = start;
2210 struct btrfs_bio_ctrl bio_ctrl = {
2211 .wbc = wbc,
2212 .opf = REQ_OP_WRITE | wbc_to_write_flags(wbc),
2213 };
2214
2215 if (wbc->no_cgroup_owner)
2216 bio_ctrl.opf |= REQ_BTRFS_CGROUP_PUNT;
2217
2218 ASSERT(IS_ALIGNED(start, sectorsize) && IS_ALIGNED(end + 1, sectorsize));
2219
2220 while (cur <= end) {
2221 u64 cur_end = min(round_down(cur, PAGE_SIZE) + PAGE_SIZE - 1, end);
2222 u32 cur_len = cur_end + 1 - cur;
2223 struct page *page;
2224 int nr = 0;
2225
2226 page = find_get_page(mapping, cur >> PAGE_SHIFT);
2227 ASSERT(PageLocked(page));
2228 if (pages_dirty && page != locked_page) {
2229 ASSERT(PageDirty(page));
2230 clear_page_dirty_for_io(page);
2231 }
2232
2233 ret = __extent_writepage_io(BTRFS_I(inode), page, &bio_ctrl,
2234 i_size, &nr);
2235 if (ret == 1)
2236 goto next_page;
2237
2238 /* Make sure the mapping tag for page dirty gets cleared. */
2239 if (nr == 0) {
2240 set_page_writeback(page);
2241 end_page_writeback(page);
2242 }
2243 if (ret) {
2244 btrfs_mark_ordered_io_finished(BTRFS_I(inode), page,
2245 cur, cur_len, !ret);
2246 mapping_set_error(page->mapping, ret);
2247 }
2248 btrfs_folio_unlock_writer(fs_info, page_folio(page), cur, cur_len);
2249 if (ret < 0)
2250 found_error = true;
2251next_page:
2252 put_page(page);
2253 cur = cur_end + 1;
2254 }
2255
2256 submit_write_bio(&bio_ctrl, found_error ? ret : 0);
2257}
2258
2259int extent_writepages(struct address_space *mapping,
2260 struct writeback_control *wbc)
2261{
2262 struct inode *inode = mapping->host;
2263 int ret = 0;
2264 struct btrfs_bio_ctrl bio_ctrl = {
2265 .wbc = wbc,
2266 .opf = REQ_OP_WRITE | wbc_to_write_flags(wbc),
2267 };
2268
2269 /*
2270 * Allow only a single thread to do the reloc work in zoned mode to
2271 * protect the write pointer updates.
2272 */
2273 btrfs_zoned_data_reloc_lock(BTRFS_I(inode));
2274 ret = extent_write_cache_pages(mapping, &bio_ctrl);
2275 submit_write_bio(&bio_ctrl, ret);
2276 btrfs_zoned_data_reloc_unlock(BTRFS_I(inode));
2277 return ret;
2278}
2279
2280void extent_readahead(struct readahead_control *rac)
2281{
2282 struct btrfs_bio_ctrl bio_ctrl = { .opf = REQ_OP_READ | REQ_RAHEAD };
2283 struct page *pagepool[16];
2284 struct extent_map *em_cached = NULL;
2285 u64 prev_em_start = (u64)-1;
2286 int nr;
2287
2288 while ((nr = readahead_page_batch(rac, pagepool))) {
2289 u64 contig_start = readahead_pos(rac);
2290 u64 contig_end = contig_start + readahead_batch_length(rac) - 1;
2291
2292 contiguous_readpages(pagepool, nr, contig_start, contig_end,
2293 &em_cached, &bio_ctrl, &prev_em_start);
2294 }
2295
2296 if (em_cached)
2297 free_extent_map(em_cached);
2298 submit_one_bio(&bio_ctrl);
2299}
2300
2301/*
2302 * basic invalidate_folio code, this waits on any locked or writeback
2303 * ranges corresponding to the folio, and then deletes any extent state
2304 * records from the tree
2305 */
2306int extent_invalidate_folio(struct extent_io_tree *tree,
2307 struct folio *folio, size_t offset)
2308{
2309 struct extent_state *cached_state = NULL;
2310 u64 start = folio_pos(folio);
2311 u64 end = start + folio_size(folio) - 1;
2312 size_t blocksize = folio->mapping->host->i_sb->s_blocksize;
2313
2314 /* This function is only called for the btree inode */
2315 ASSERT(tree->owner == IO_TREE_BTREE_INODE_IO);
2316
2317 start += ALIGN(offset, blocksize);
2318 if (start > end)
2319 return 0;
2320
2321 lock_extent(tree, start, end, &cached_state);
2322 folio_wait_writeback(folio);
2323
2324 /*
2325 * Currently for btree io tree, only EXTENT_LOCKED is utilized,
2326 * so here we only need to unlock the extent range to free any
2327 * existing extent state.
2328 */
2329 unlock_extent(tree, start, end, &cached_state);
2330 return 0;
2331}
2332
2333/*
2334 * a helper for release_folio, this tests for areas of the page that
2335 * are locked or under IO and drops the related state bits if it is safe
2336 * to drop the page.
2337 */
2338static int try_release_extent_state(struct extent_io_tree *tree,
2339 struct page *page, gfp_t mask)
2340{
2341 u64 start = page_offset(page);
2342 u64 end = start + PAGE_SIZE - 1;
2343 int ret = 1;
2344
2345 if (test_range_bit_exists(tree, start, end, EXTENT_LOCKED)) {
2346 ret = 0;
2347 } else {
2348 u32 clear_bits = ~(EXTENT_LOCKED | EXTENT_NODATASUM |
2349 EXTENT_DELALLOC_NEW | EXTENT_CTLBITS |
2350 EXTENT_QGROUP_RESERVED);
2351
2352 /*
2353 * At this point we can safely clear everything except the
2354 * locked bit, the nodatasum bit and the delalloc new bit.
2355 * The delalloc new bit will be cleared by ordered extent
2356 * completion.
2357 */
2358 ret = __clear_extent_bit(tree, start, end, clear_bits, NULL, NULL);
2359
2360 /* if clear_extent_bit failed for enomem reasons,
2361 * we can't allow the release to continue.
2362 */
2363 if (ret < 0)
2364 ret = 0;
2365 else
2366 ret = 1;
2367 }
2368 return ret;
2369}
2370
2371/*
2372 * a helper for release_folio. As long as there are no locked extents
2373 * in the range corresponding to the page, both state records and extent
2374 * map records are removed
2375 */
2376int try_release_extent_mapping(struct page *page, gfp_t mask)
2377{
2378 struct extent_map *em;
2379 u64 start = page_offset(page);
2380 u64 end = start + PAGE_SIZE - 1;
2381 struct btrfs_inode *btrfs_inode = BTRFS_I(page->mapping->host);
2382 struct extent_io_tree *tree = &btrfs_inode->io_tree;
2383 struct extent_map_tree *map = &btrfs_inode->extent_tree;
2384
2385 if (gfpflags_allow_blocking(mask) &&
2386 page->mapping->host->i_size > SZ_16M) {
2387 u64 len;
2388 while (start <= end) {
2389 struct btrfs_fs_info *fs_info;
2390 u64 cur_gen;
2391
2392 len = end - start + 1;
2393 write_lock(&map->lock);
2394 em = lookup_extent_mapping(map, start, len);
2395 if (!em) {
2396 write_unlock(&map->lock);
2397 break;
2398 }
2399 if ((em->flags & EXTENT_FLAG_PINNED) ||
2400 em->start != start) {
2401 write_unlock(&map->lock);
2402 free_extent_map(em);
2403 break;
2404 }
2405 if (test_range_bit_exists(tree, em->start,
2406 extent_map_end(em) - 1,
2407 EXTENT_LOCKED))
2408 goto next;
2409 /*
2410 * If it's not in the list of modified extents, used
2411 * by a fast fsync, we can remove it. If it's being
2412 * logged we can safely remove it since fsync took an
2413 * extra reference on the em.
2414 */
2415 if (list_empty(&em->list) ||
2416 (em->flags & EXTENT_FLAG_LOGGING))
2417 goto remove_em;
2418 /*
2419 * If it's in the list of modified extents, remove it
2420 * only if its generation is older then the current one,
2421 * in which case we don't need it for a fast fsync.
2422 * Otherwise don't remove it, we could be racing with an
2423 * ongoing fast fsync that could miss the new extent.
2424 */
2425 fs_info = btrfs_inode->root->fs_info;
2426 spin_lock(&fs_info->trans_lock);
2427 cur_gen = fs_info->generation;
2428 spin_unlock(&fs_info->trans_lock);
2429 if (em->generation >= cur_gen)
2430 goto next;
2431remove_em:
2432 /*
2433 * We only remove extent maps that are not in the list of
2434 * modified extents or that are in the list but with a
2435 * generation lower then the current generation, so there
2436 * is no need to set the full fsync flag on the inode (it
2437 * hurts the fsync performance for workloads with a data
2438 * size that exceeds or is close to the system's memory).
2439 */
2440 remove_extent_mapping(map, em);
2441 /* once for the rb tree */
2442 free_extent_map(em);
2443next:
2444 start = extent_map_end(em);
2445 write_unlock(&map->lock);
2446
2447 /* once for us */
2448 free_extent_map(em);
2449
2450 cond_resched(); /* Allow large-extent preemption. */
2451 }
2452 }
2453 return try_release_extent_state(tree, page, mask);
2454}
2455
2456/*
2457 * To cache previous fiemap extent
2458 *
2459 * Will be used for merging fiemap extent
2460 */
2461struct fiemap_cache {
2462 u64 offset;
2463 u64 phys;
2464 u64 len;
2465 u32 flags;
2466 bool cached;
2467};
2468
2469/*
2470 * Helper to submit fiemap extent.
2471 *
2472 * Will try to merge current fiemap extent specified by @offset, @phys,
2473 * @len and @flags with cached one.
2474 * And only when we fails to merge, cached one will be submitted as
2475 * fiemap extent.
2476 *
2477 * Return value is the same as fiemap_fill_next_extent().
2478 */
2479static int emit_fiemap_extent(struct fiemap_extent_info *fieinfo,
2480 struct fiemap_cache *cache,
2481 u64 offset, u64 phys, u64 len, u32 flags)
2482{
2483 u64 cache_end;
2484 int ret = 0;
2485
2486 /* Set at the end of extent_fiemap(). */
2487 ASSERT((flags & FIEMAP_EXTENT_LAST) == 0);
2488
2489 if (!cache->cached)
2490 goto assign;
2491
2492 /*
2493 * When iterating the extents of the inode, at extent_fiemap(), we may
2494 * find an extent that starts at an offset behind the end offset of the
2495 * previous extent we processed. This happens if fiemap is called
2496 * without FIEMAP_FLAG_SYNC and there are ordered extents completing
2497 * while we call btrfs_next_leaf() (through fiemap_next_leaf_item()).
2498 *
2499 * For example we are in leaf X processing its last item, which is the
2500 * file extent item for file range [512K, 1M[, and after
2501 * btrfs_next_leaf() releases the path, there's an ordered extent that
2502 * completes for the file range [768K, 2M[, and that results in trimming
2503 * the file extent item so that it now corresponds to the file range
2504 * [512K, 768K[ and a new file extent item is inserted for the file
2505 * range [768K, 2M[, which may end up as the last item of leaf X or as
2506 * the first item of the next leaf - in either case btrfs_next_leaf()
2507 * will leave us with a path pointing to the new extent item, for the
2508 * file range [768K, 2M[, since that's the first key that follows the
2509 * last one we processed. So in order not to report overlapping extents
2510 * to user space, we trim the length of the previously cached extent and
2511 * emit it.
2512 *
2513 * Upon calling btrfs_next_leaf() we may also find an extent with an
2514 * offset smaller than or equals to cache->offset, and this happens
2515 * when we had a hole or prealloc extent with several delalloc ranges in
2516 * it, but after btrfs_next_leaf() released the path, delalloc was
2517 * flushed and the resulting ordered extents were completed, so we can
2518 * now have found a file extent item for an offset that is smaller than
2519 * or equals to what we have in cache->offset. We deal with this as
2520 * described below.
2521 */
2522 cache_end = cache->offset + cache->len;
2523 if (cache_end > offset) {
2524 if (offset == cache->offset) {
2525 /*
2526 * We cached a dealloc range (found in the io tree) for
2527 * a hole or prealloc extent and we have now found a
2528 * file extent item for the same offset. What we have
2529 * now is more recent and up to date, so discard what
2530 * we had in the cache and use what we have just found.
2531 */
2532 goto assign;
2533 } else if (offset > cache->offset) {
2534 /*
2535 * The extent range we previously found ends after the
2536 * offset of the file extent item we found and that
2537 * offset falls somewhere in the middle of that previous
2538 * extent range. So adjust the range we previously found
2539 * to end at the offset of the file extent item we have
2540 * just found, since this extent is more up to date.
2541 * Emit that adjusted range and cache the file extent
2542 * item we have just found. This corresponds to the case
2543 * where a previously found file extent item was split
2544 * due to an ordered extent completing.
2545 */
2546 cache->len = offset - cache->offset;
2547 goto emit;
2548 } else {
2549 const u64 range_end = offset + len;
2550
2551 /*
2552 * The offset of the file extent item we have just found
2553 * is behind the cached offset. This means we were
2554 * processing a hole or prealloc extent for which we
2555 * have found delalloc ranges (in the io tree), so what
2556 * we have in the cache is the last delalloc range we
2557 * found while the file extent item we found can be
2558 * either for a whole delalloc range we previously
2559 * emmitted or only a part of that range.
2560 *
2561 * We have two cases here:
2562 *
2563 * 1) The file extent item's range ends at or behind the
2564 * cached extent's end. In this case just ignore the
2565 * current file extent item because we don't want to
2566 * overlap with previous ranges that may have been
2567 * emmitted already;
2568 *
2569 * 2) The file extent item starts behind the currently
2570 * cached extent but its end offset goes beyond the
2571 * end offset of the cached extent. We don't want to
2572 * overlap with a previous range that may have been
2573 * emmitted already, so we emit the currently cached
2574 * extent and then partially store the current file
2575 * extent item's range in the cache, for the subrange
2576 * going the cached extent's end to the end of the
2577 * file extent item.
2578 */
2579 if (range_end <= cache_end)
2580 return 0;
2581
2582 if (!(flags & (FIEMAP_EXTENT_ENCODED | FIEMAP_EXTENT_DELALLOC)))
2583 phys += cache_end - offset;
2584
2585 offset = cache_end;
2586 len = range_end - cache_end;
2587 goto emit;
2588 }
2589 }
2590
2591 /*
2592 * Only merges fiemap extents if
2593 * 1) Their logical addresses are continuous
2594 *
2595 * 2) Their physical addresses are continuous
2596 * So truly compressed (physical size smaller than logical size)
2597 * extents won't get merged with each other
2598 *
2599 * 3) Share same flags
2600 */
2601 if (cache->offset + cache->len == offset &&
2602 cache->phys + cache->len == phys &&
2603 cache->flags == flags) {
2604 cache->len += len;
2605 return 0;
2606 }
2607
2608emit:
2609 /* Not mergeable, need to submit cached one */
2610 ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
2611 cache->len, cache->flags);
2612 cache->cached = false;
2613 if (ret)
2614 return ret;
2615assign:
2616 cache->cached = true;
2617 cache->offset = offset;
2618 cache->phys = phys;
2619 cache->len = len;
2620 cache->flags = flags;
2621
2622 return 0;
2623}
2624
2625/*
2626 * Emit last fiemap cache
2627 *
2628 * The last fiemap cache may still be cached in the following case:
2629 * 0 4k 8k
2630 * |<- Fiemap range ->|
2631 * |<------------ First extent ----------->|
2632 *
2633 * In this case, the first extent range will be cached but not emitted.
2634 * So we must emit it before ending extent_fiemap().
2635 */
2636static int emit_last_fiemap_cache(struct fiemap_extent_info *fieinfo,
2637 struct fiemap_cache *cache)
2638{
2639 int ret;
2640
2641 if (!cache->cached)
2642 return 0;
2643
2644 ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
2645 cache->len, cache->flags);
2646 cache->cached = false;
2647 if (ret > 0)
2648 ret = 0;
2649 return ret;
2650}
2651
2652static int fiemap_next_leaf_item(struct btrfs_inode *inode, struct btrfs_path *path)
2653{
2654 struct extent_buffer *clone;
2655 struct btrfs_key key;
2656 int slot;
2657 int ret;
2658
2659 path->slots[0]++;
2660 if (path->slots[0] < btrfs_header_nritems(path->nodes[0]))
2661 return 0;
2662
2663 ret = btrfs_next_leaf(inode->root, path);
2664 if (ret != 0)
2665 return ret;
2666
2667 /*
2668 * Don't bother with cloning if there are no more file extent items for
2669 * our inode.
2670 */
2671 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2672 if (key.objectid != btrfs_ino(inode) || key.type != BTRFS_EXTENT_DATA_KEY)
2673 return 1;
2674
2675 /* See the comment at fiemap_search_slot() about why we clone. */
2676 clone = btrfs_clone_extent_buffer(path->nodes[0]);
2677 if (!clone)
2678 return -ENOMEM;
2679
2680 slot = path->slots[0];
2681 btrfs_release_path(path);
2682 path->nodes[0] = clone;
2683 path->slots[0] = slot;
2684
2685 return 0;
2686}
2687
2688/*
2689 * Search for the first file extent item that starts at a given file offset or
2690 * the one that starts immediately before that offset.
2691 * Returns: 0 on success, < 0 on error, 1 if not found.
2692 */
2693static int fiemap_search_slot(struct btrfs_inode *inode, struct btrfs_path *path,
2694 u64 file_offset)
2695{
2696 const u64 ino = btrfs_ino(inode);
2697 struct btrfs_root *root = inode->root;
2698 struct extent_buffer *clone;
2699 struct btrfs_key key;
2700 int slot;
2701 int ret;
2702
2703 key.objectid = ino;
2704 key.type = BTRFS_EXTENT_DATA_KEY;
2705 key.offset = file_offset;
2706
2707 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2708 if (ret < 0)
2709 return ret;
2710
2711 if (ret > 0 && path->slots[0] > 0) {
2712 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
2713 if (key.objectid == ino && key.type == BTRFS_EXTENT_DATA_KEY)
2714 path->slots[0]--;
2715 }
2716
2717 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
2718 ret = btrfs_next_leaf(root, path);
2719 if (ret != 0)
2720 return ret;
2721
2722 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2723 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
2724 return 1;
2725 }
2726
2727 /*
2728 * We clone the leaf and use it during fiemap. This is because while
2729 * using the leaf we do expensive things like checking if an extent is
2730 * shared, which can take a long time. In order to prevent blocking
2731 * other tasks for too long, we use a clone of the leaf. We have locked
2732 * the file range in the inode's io tree, so we know none of our file
2733 * extent items can change. This way we avoid blocking other tasks that
2734 * want to insert items for other inodes in the same leaf or b+tree
2735 * rebalance operations (triggered for example when someone is trying
2736 * to push items into this leaf when trying to insert an item in a
2737 * neighbour leaf).
2738 * We also need the private clone because holding a read lock on an
2739 * extent buffer of the subvolume's b+tree will make lockdep unhappy
2740 * when we call fiemap_fill_next_extent(), because that may cause a page
2741 * fault when filling the user space buffer with fiemap data.
2742 */
2743 clone = btrfs_clone_extent_buffer(path->nodes[0]);
2744 if (!clone)
2745 return -ENOMEM;
2746
2747 slot = path->slots[0];
2748 btrfs_release_path(path);
2749 path->nodes[0] = clone;
2750 path->slots[0] = slot;
2751
2752 return 0;
2753}
2754
2755/*
2756 * Process a range which is a hole or a prealloc extent in the inode's subvolume
2757 * btree. If @disk_bytenr is 0, we are dealing with a hole, otherwise a prealloc
2758 * extent. The end offset (@end) is inclusive.
2759 */
2760static int fiemap_process_hole(struct btrfs_inode *inode,
2761 struct fiemap_extent_info *fieinfo,
2762 struct fiemap_cache *cache,
2763 struct extent_state **delalloc_cached_state,
2764 struct btrfs_backref_share_check_ctx *backref_ctx,
2765 u64 disk_bytenr, u64 extent_offset,
2766 u64 extent_gen,
2767 u64 start, u64 end)
2768{
2769 const u64 i_size = i_size_read(&inode->vfs_inode);
2770 u64 cur_offset = start;
2771 u64 last_delalloc_end = 0;
2772 u32 prealloc_flags = FIEMAP_EXTENT_UNWRITTEN;
2773 bool checked_extent_shared = false;
2774 int ret;
2775
2776 /*
2777 * There can be no delalloc past i_size, so don't waste time looking for
2778 * it beyond i_size.
2779 */
2780 while (cur_offset < end && cur_offset < i_size) {
2781 struct extent_state *cached_state = NULL;
2782 u64 delalloc_start;
2783 u64 delalloc_end;
2784 u64 prealloc_start;
2785 u64 lockstart;
2786 u64 lockend;
2787 u64 prealloc_len = 0;
2788 bool delalloc;
2789
2790 lockstart = round_down(cur_offset, inode->root->fs_info->sectorsize);
2791 lockend = round_up(end, inode->root->fs_info->sectorsize);
2792
2793 /*
2794 * We are only locking for the delalloc range because that's the
2795 * only thing that can change here. With fiemap we have a lock
2796 * on the inode, so no buffered or direct writes can happen.
2797 *
2798 * However mmaps and normal page writeback will cause this to
2799 * change arbitrarily. We have to lock the extent lock here to
2800 * make sure that nobody messes with the tree while we're doing
2801 * btrfs_find_delalloc_in_range.
2802 */
2803 lock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
2804 delalloc = btrfs_find_delalloc_in_range(inode, cur_offset, end,
2805 delalloc_cached_state,
2806 &delalloc_start,
2807 &delalloc_end);
2808 unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
2809 if (!delalloc)
2810 break;
2811
2812 /*
2813 * If this is a prealloc extent we have to report every section
2814 * of it that has no delalloc.
2815 */
2816 if (disk_bytenr != 0) {
2817 if (last_delalloc_end == 0) {
2818 prealloc_start = start;
2819 prealloc_len = delalloc_start - start;
2820 } else {
2821 prealloc_start = last_delalloc_end + 1;
2822 prealloc_len = delalloc_start - prealloc_start;
2823 }
2824 }
2825
2826 if (prealloc_len > 0) {
2827 if (!checked_extent_shared && fieinfo->fi_extents_max) {
2828 ret = btrfs_is_data_extent_shared(inode,
2829 disk_bytenr,
2830 extent_gen,
2831 backref_ctx);
2832 if (ret < 0)
2833 return ret;
2834 else if (ret > 0)
2835 prealloc_flags |= FIEMAP_EXTENT_SHARED;
2836
2837 checked_extent_shared = true;
2838 }
2839 ret = emit_fiemap_extent(fieinfo, cache, prealloc_start,
2840 disk_bytenr + extent_offset,
2841 prealloc_len, prealloc_flags);
2842 if (ret)
2843 return ret;
2844 extent_offset += prealloc_len;
2845 }
2846
2847 ret = emit_fiemap_extent(fieinfo, cache, delalloc_start, 0,
2848 delalloc_end + 1 - delalloc_start,
2849 FIEMAP_EXTENT_DELALLOC |
2850 FIEMAP_EXTENT_UNKNOWN);
2851 if (ret)
2852 return ret;
2853
2854 last_delalloc_end = delalloc_end;
2855 cur_offset = delalloc_end + 1;
2856 extent_offset += cur_offset - delalloc_start;
2857 cond_resched();
2858 }
2859
2860 /*
2861 * Either we found no delalloc for the whole prealloc extent or we have
2862 * a prealloc extent that spans i_size or starts at or after i_size.
2863 */
2864 if (disk_bytenr != 0 && last_delalloc_end < end) {
2865 u64 prealloc_start;
2866 u64 prealloc_len;
2867
2868 if (last_delalloc_end == 0) {
2869 prealloc_start = start;
2870 prealloc_len = end + 1 - start;
2871 } else {
2872 prealloc_start = last_delalloc_end + 1;
2873 prealloc_len = end + 1 - prealloc_start;
2874 }
2875
2876 if (!checked_extent_shared && fieinfo->fi_extents_max) {
2877 ret = btrfs_is_data_extent_shared(inode,
2878 disk_bytenr,
2879 extent_gen,
2880 backref_ctx);
2881 if (ret < 0)
2882 return ret;
2883 else if (ret > 0)
2884 prealloc_flags |= FIEMAP_EXTENT_SHARED;
2885 }
2886 ret = emit_fiemap_extent(fieinfo, cache, prealloc_start,
2887 disk_bytenr + extent_offset,
2888 prealloc_len, prealloc_flags);
2889 if (ret)
2890 return ret;
2891 }
2892
2893 return 0;
2894}
2895
2896static int fiemap_find_last_extent_offset(struct btrfs_inode *inode,
2897 struct btrfs_path *path,
2898 u64 *last_extent_end_ret)
2899{
2900 const u64 ino = btrfs_ino(inode);
2901 struct btrfs_root *root = inode->root;
2902 struct extent_buffer *leaf;
2903 struct btrfs_file_extent_item *ei;
2904 struct btrfs_key key;
2905 u64 disk_bytenr;
2906 int ret;
2907
2908 /*
2909 * Lookup the last file extent. We're not using i_size here because
2910 * there might be preallocation past i_size.
2911 */
2912 ret = btrfs_lookup_file_extent(NULL, root, path, ino, (u64)-1, 0);
2913 /* There can't be a file extent item at offset (u64)-1 */
2914 ASSERT(ret != 0);
2915 if (ret < 0)
2916 return ret;
2917
2918 /*
2919 * For a non-existing key, btrfs_search_slot() always leaves us at a
2920 * slot > 0, except if the btree is empty, which is impossible because
2921 * at least it has the inode item for this inode and all the items for
2922 * the root inode 256.
2923 */
2924 ASSERT(path->slots[0] > 0);
2925 path->slots[0]--;
2926 leaf = path->nodes[0];
2927 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2928 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY) {
2929 /* No file extent items in the subvolume tree. */
2930 *last_extent_end_ret = 0;
2931 return 0;
2932 }
2933
2934 /*
2935 * For an inline extent, the disk_bytenr is where inline data starts at,
2936 * so first check if we have an inline extent item before checking if we
2937 * have an implicit hole (disk_bytenr == 0).
2938 */
2939 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_file_extent_item);
2940 if (btrfs_file_extent_type(leaf, ei) == BTRFS_FILE_EXTENT_INLINE) {
2941 *last_extent_end_ret = btrfs_file_extent_end(path);
2942 return 0;
2943 }
2944
2945 /*
2946 * Find the last file extent item that is not a hole (when NO_HOLES is
2947 * not enabled). This should take at most 2 iterations in the worst
2948 * case: we have one hole file extent item at slot 0 of a leaf and
2949 * another hole file extent item as the last item in the previous leaf.
2950 * This is because we merge file extent items that represent holes.
2951 */
2952 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, ei);
2953 while (disk_bytenr == 0) {
2954 ret = btrfs_previous_item(root, path, ino, BTRFS_EXTENT_DATA_KEY);
2955 if (ret < 0) {
2956 return ret;
2957 } else if (ret > 0) {
2958 /* No file extent items that are not holes. */
2959 *last_extent_end_ret = 0;
2960 return 0;
2961 }
2962 leaf = path->nodes[0];
2963 ei = btrfs_item_ptr(leaf, path->slots[0],
2964 struct btrfs_file_extent_item);
2965 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, ei);
2966 }
2967
2968 *last_extent_end_ret = btrfs_file_extent_end(path);
2969 return 0;
2970}
2971
2972int extent_fiemap(struct btrfs_inode *inode, struct fiemap_extent_info *fieinfo,
2973 u64 start, u64 len)
2974{
2975 const u64 ino = btrfs_ino(inode);
2976 struct extent_state *delalloc_cached_state = NULL;
2977 struct btrfs_path *path;
2978 struct fiemap_cache cache = { 0 };
2979 struct btrfs_backref_share_check_ctx *backref_ctx;
2980 u64 last_extent_end;
2981 u64 prev_extent_end;
2982 u64 range_start;
2983 u64 range_end;
2984 const u64 sectorsize = inode->root->fs_info->sectorsize;
2985 bool stopped = false;
2986 int ret;
2987
2988 backref_ctx = btrfs_alloc_backref_share_check_ctx();
2989 path = btrfs_alloc_path();
2990 if (!backref_ctx || !path) {
2991 ret = -ENOMEM;
2992 goto out;
2993 }
2994
2995 range_start = round_down(start, sectorsize);
2996 range_end = round_up(start + len, sectorsize);
2997 prev_extent_end = range_start;
2998
2999 ret = fiemap_find_last_extent_offset(inode, path, &last_extent_end);
3000 if (ret < 0)
3001 goto out;
3002 btrfs_release_path(path);
3003
3004 path->reada = READA_FORWARD;
3005 ret = fiemap_search_slot(inode, path, range_start);
3006 if (ret < 0) {
3007 goto out;
3008 } else if (ret > 0) {
3009 /*
3010 * No file extent item found, but we may have delalloc between
3011 * the current offset and i_size. So check for that.
3012 */
3013 ret = 0;
3014 goto check_eof_delalloc;
3015 }
3016
3017 while (prev_extent_end < range_end) {
3018 struct extent_buffer *leaf = path->nodes[0];
3019 struct btrfs_file_extent_item *ei;
3020 struct btrfs_key key;
3021 u64 extent_end;
3022 u64 extent_len;
3023 u64 extent_offset = 0;
3024 u64 extent_gen;
3025 u64 disk_bytenr = 0;
3026 u64 flags = 0;
3027 int extent_type;
3028 u8 compression;
3029
3030 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3031 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
3032 break;
3033
3034 extent_end = btrfs_file_extent_end(path);
3035
3036 /*
3037 * The first iteration can leave us at an extent item that ends
3038 * before our range's start. Move to the next item.
3039 */
3040 if (extent_end <= range_start)
3041 goto next_item;
3042
3043 backref_ctx->curr_leaf_bytenr = leaf->start;
3044
3045 /* We have in implicit hole (NO_HOLES feature enabled). */
3046 if (prev_extent_end < key.offset) {
3047 const u64 hole_end = min(key.offset, range_end) - 1;
3048
3049 ret = fiemap_process_hole(inode, fieinfo, &cache,
3050 &delalloc_cached_state,
3051 backref_ctx, 0, 0, 0,
3052 prev_extent_end, hole_end);
3053 if (ret < 0) {
3054 goto out;
3055 } else if (ret > 0) {
3056 /* fiemap_fill_next_extent() told us to stop. */
3057 stopped = true;
3058 break;
3059 }
3060
3061 /* We've reached the end of the fiemap range, stop. */
3062 if (key.offset >= range_end) {
3063 stopped = true;
3064 break;
3065 }
3066 }
3067
3068 extent_len = extent_end - key.offset;
3069 ei = btrfs_item_ptr(leaf, path->slots[0],
3070 struct btrfs_file_extent_item);
3071 compression = btrfs_file_extent_compression(leaf, ei);
3072 extent_type = btrfs_file_extent_type(leaf, ei);
3073 extent_gen = btrfs_file_extent_generation(leaf, ei);
3074
3075 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
3076 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, ei);
3077 if (compression == BTRFS_COMPRESS_NONE)
3078 extent_offset = btrfs_file_extent_offset(leaf, ei);
3079 }
3080
3081 if (compression != BTRFS_COMPRESS_NONE)
3082 flags |= FIEMAP_EXTENT_ENCODED;
3083
3084 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
3085 flags |= FIEMAP_EXTENT_DATA_INLINE;
3086 flags |= FIEMAP_EXTENT_NOT_ALIGNED;
3087 ret = emit_fiemap_extent(fieinfo, &cache, key.offset, 0,
3088 extent_len, flags);
3089 } else if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
3090 ret = fiemap_process_hole(inode, fieinfo, &cache,
3091 &delalloc_cached_state,
3092 backref_ctx,
3093 disk_bytenr, extent_offset,
3094 extent_gen, key.offset,
3095 extent_end - 1);
3096 } else if (disk_bytenr == 0) {
3097 /* We have an explicit hole. */
3098 ret = fiemap_process_hole(inode, fieinfo, &cache,
3099 &delalloc_cached_state,
3100 backref_ctx, 0, 0, 0,
3101 key.offset, extent_end - 1);
3102 } else {
3103 /* We have a regular extent. */
3104 if (fieinfo->fi_extents_max) {
3105 ret = btrfs_is_data_extent_shared(inode,
3106 disk_bytenr,
3107 extent_gen,
3108 backref_ctx);
3109 if (ret < 0)
3110 goto out;
3111 else if (ret > 0)
3112 flags |= FIEMAP_EXTENT_SHARED;
3113 }
3114
3115 ret = emit_fiemap_extent(fieinfo, &cache, key.offset,
3116 disk_bytenr + extent_offset,
3117 extent_len, flags);
3118 }
3119
3120 if (ret < 0) {
3121 goto out;
3122 } else if (ret > 0) {
3123 /* fiemap_fill_next_extent() told us to stop. */
3124 stopped = true;
3125 break;
3126 }
3127
3128 prev_extent_end = extent_end;
3129next_item:
3130 if (fatal_signal_pending(current)) {
3131 ret = -EINTR;
3132 goto out;
3133 }
3134
3135 ret = fiemap_next_leaf_item(inode, path);
3136 if (ret < 0) {
3137 goto out;
3138 } else if (ret > 0) {
3139 /* No more file extent items for this inode. */
3140 break;
3141 }
3142 cond_resched();
3143 }
3144
3145check_eof_delalloc:
3146 /*
3147 * Release (and free) the path before emitting any final entries to
3148 * fiemap_fill_next_extent() to keep lockdep happy. This is because
3149 * once we find no more file extent items exist, we may have a
3150 * non-cloned leaf, and fiemap_fill_next_extent() can trigger page
3151 * faults when copying data to the user space buffer.
3152 */
3153 btrfs_free_path(path);
3154 path = NULL;
3155
3156 if (!stopped && prev_extent_end < range_end) {
3157 ret = fiemap_process_hole(inode, fieinfo, &cache,
3158 &delalloc_cached_state, backref_ctx,
3159 0, 0, 0, prev_extent_end, range_end - 1);
3160 if (ret < 0)
3161 goto out;
3162 prev_extent_end = range_end;
3163 }
3164
3165 if (cache.cached && cache.offset + cache.len >= last_extent_end) {
3166 const u64 i_size = i_size_read(&inode->vfs_inode);
3167
3168 if (prev_extent_end < i_size) {
3169 struct extent_state *cached_state = NULL;
3170 u64 delalloc_start;
3171 u64 delalloc_end;
3172 u64 lockstart;
3173 u64 lockend;
3174 bool delalloc;
3175
3176 lockstart = round_down(prev_extent_end, sectorsize);
3177 lockend = round_up(i_size, sectorsize);
3178
3179 /*
3180 * See the comment in fiemap_process_hole as to why
3181 * we're doing the locking here.
3182 */
3183 lock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3184 delalloc = btrfs_find_delalloc_in_range(inode,
3185 prev_extent_end,
3186 i_size - 1,
3187 &delalloc_cached_state,
3188 &delalloc_start,
3189 &delalloc_end);
3190 unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3191 if (!delalloc)
3192 cache.flags |= FIEMAP_EXTENT_LAST;
3193 } else {
3194 cache.flags |= FIEMAP_EXTENT_LAST;
3195 }
3196 }
3197
3198 ret = emit_last_fiemap_cache(fieinfo, &cache);
3199out:
3200 free_extent_state(delalloc_cached_state);
3201 btrfs_free_backref_share_ctx(backref_ctx);
3202 btrfs_free_path(path);
3203 return ret;
3204}
3205
3206static void __free_extent_buffer(struct extent_buffer *eb)
3207{
3208 kmem_cache_free(extent_buffer_cache, eb);
3209}
3210
3211static int extent_buffer_under_io(const struct extent_buffer *eb)
3212{
3213 return (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
3214 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
3215}
3216
3217static bool folio_range_has_eb(struct btrfs_fs_info *fs_info, struct folio *folio)
3218{
3219 struct btrfs_subpage *subpage;
3220
3221 lockdep_assert_held(&folio->mapping->i_private_lock);
3222
3223 if (folio_test_private(folio)) {
3224 subpage = folio_get_private(folio);
3225 if (atomic_read(&subpage->eb_refs))
3226 return true;
3227 /*
3228 * Even there is no eb refs here, we may still have
3229 * end_page_read() call relying on page::private.
3230 */
3231 if (atomic_read(&subpage->readers))
3232 return true;
3233 }
3234 return false;
3235}
3236
3237static void detach_extent_buffer_folio(struct extent_buffer *eb, struct folio *folio)
3238{
3239 struct btrfs_fs_info *fs_info = eb->fs_info;
3240 const bool mapped = !test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
3241
3242 /*
3243 * For mapped eb, we're going to change the folio private, which should
3244 * be done under the i_private_lock.
3245 */
3246 if (mapped)
3247 spin_lock(&folio->mapping->i_private_lock);
3248
3249 if (!folio_test_private(folio)) {
3250 if (mapped)
3251 spin_unlock(&folio->mapping->i_private_lock);
3252 return;
3253 }
3254
3255 if (fs_info->nodesize >= PAGE_SIZE) {
3256 /*
3257 * We do this since we'll remove the pages after we've
3258 * removed the eb from the radix tree, so we could race
3259 * and have this page now attached to the new eb. So
3260 * only clear folio if it's still connected to
3261 * this eb.
3262 */
3263 if (folio_test_private(folio) && folio_get_private(folio) == eb) {
3264 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
3265 BUG_ON(folio_test_dirty(folio));
3266 BUG_ON(folio_test_writeback(folio));
3267 /* We need to make sure we haven't be attached to a new eb. */
3268 folio_detach_private(folio);
3269 }
3270 if (mapped)
3271 spin_unlock(&folio->mapping->i_private_lock);
3272 return;
3273 }
3274
3275 /*
3276 * For subpage, we can have dummy eb with folio private attached. In
3277 * this case, we can directly detach the private as such folio is only
3278 * attached to one dummy eb, no sharing.
3279 */
3280 if (!mapped) {
3281 btrfs_detach_subpage(fs_info, folio);
3282 return;
3283 }
3284
3285 btrfs_folio_dec_eb_refs(fs_info, folio);
3286
3287 /*
3288 * We can only detach the folio private if there are no other ebs in the
3289 * page range and no unfinished IO.
3290 */
3291 if (!folio_range_has_eb(fs_info, folio))
3292 btrfs_detach_subpage(fs_info, folio);
3293
3294 spin_unlock(&folio->mapping->i_private_lock);
3295}
3296
3297/* Release all pages attached to the extent buffer */
3298static void btrfs_release_extent_buffer_pages(struct extent_buffer *eb)
3299{
3300 ASSERT(!extent_buffer_under_io(eb));
3301
3302 for (int i = 0; i < INLINE_EXTENT_BUFFER_PAGES; i++) {
3303 struct folio *folio = eb->folios[i];
3304
3305 if (!folio)
3306 continue;
3307
3308 detach_extent_buffer_folio(eb, folio);
3309
3310 /* One for when we allocated the folio. */
3311 folio_put(folio);
3312 }
3313}
3314
3315/*
3316 * Helper for releasing the extent buffer.
3317 */
3318static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
3319{
3320 btrfs_release_extent_buffer_pages(eb);
3321 btrfs_leak_debug_del_eb(eb);
3322 __free_extent_buffer(eb);
3323}
3324
3325static struct extent_buffer *
3326__alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
3327 unsigned long len)
3328{
3329 struct extent_buffer *eb = NULL;
3330
3331 eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
3332 eb->start = start;
3333 eb->len = len;
3334 eb->fs_info = fs_info;
3335 init_rwsem(&eb->lock);
3336
3337 btrfs_leak_debug_add_eb(eb);
3338
3339 spin_lock_init(&eb->refs_lock);
3340 atomic_set(&eb->refs, 1);
3341
3342 ASSERT(len <= BTRFS_MAX_METADATA_BLOCKSIZE);
3343
3344 return eb;
3345}
3346
3347struct extent_buffer *btrfs_clone_extent_buffer(const struct extent_buffer *src)
3348{
3349 struct extent_buffer *new;
3350 int num_folios = num_extent_folios(src);
3351 int ret;
3352
3353 new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
3354 if (new == NULL)
3355 return NULL;
3356
3357 /*
3358 * Set UNMAPPED before calling btrfs_release_extent_buffer(), as
3359 * btrfs_release_extent_buffer() have different behavior for
3360 * UNMAPPED subpage extent buffer.
3361 */
3362 set_bit(EXTENT_BUFFER_UNMAPPED, &new->bflags);
3363
3364 ret = alloc_eb_folio_array(new, 0);
3365 if (ret) {
3366 btrfs_release_extent_buffer(new);
3367 return NULL;
3368 }
3369
3370 for (int i = 0; i < num_folios; i++) {
3371 struct folio *folio = new->folios[i];
3372 int ret;
3373
3374 ret = attach_extent_buffer_folio(new, folio, NULL);
3375 if (ret < 0) {
3376 btrfs_release_extent_buffer(new);
3377 return NULL;
3378 }
3379 WARN_ON(folio_test_dirty(folio));
3380 }
3381 copy_extent_buffer_full(new, src);
3382 set_extent_buffer_uptodate(new);
3383
3384 return new;
3385}
3386
3387struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
3388 u64 start, unsigned long len)
3389{
3390 struct extent_buffer *eb;
3391 int num_folios = 0;
3392 int ret;
3393
3394 eb = __alloc_extent_buffer(fs_info, start, len);
3395 if (!eb)
3396 return NULL;
3397
3398 ret = alloc_eb_folio_array(eb, 0);
3399 if (ret)
3400 goto err;
3401
3402 num_folios = num_extent_folios(eb);
3403 for (int i = 0; i < num_folios; i++) {
3404 ret = attach_extent_buffer_folio(eb, eb->folios[i], NULL);
3405 if (ret < 0)
3406 goto err;
3407 }
3408
3409 set_extent_buffer_uptodate(eb);
3410 btrfs_set_header_nritems(eb, 0);
3411 set_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
3412
3413 return eb;
3414err:
3415 for (int i = 0; i < num_folios; i++) {
3416 if (eb->folios[i]) {
3417 detach_extent_buffer_folio(eb, eb->folios[i]);
3418 __folio_put(eb->folios[i]);
3419 }
3420 }
3421 __free_extent_buffer(eb);
3422 return NULL;
3423}
3424
3425struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
3426 u64 start)
3427{
3428 return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
3429}
3430
3431static void check_buffer_tree_ref(struct extent_buffer *eb)
3432{
3433 int refs;
3434 /*
3435 * The TREE_REF bit is first set when the extent_buffer is added
3436 * to the radix tree. It is also reset, if unset, when a new reference
3437 * is created by find_extent_buffer.
3438 *
3439 * It is only cleared in two cases: freeing the last non-tree
3440 * reference to the extent_buffer when its STALE bit is set or
3441 * calling release_folio when the tree reference is the only reference.
3442 *
3443 * In both cases, care is taken to ensure that the extent_buffer's
3444 * pages are not under io. However, release_folio can be concurrently
3445 * called with creating new references, which is prone to race
3446 * conditions between the calls to check_buffer_tree_ref in those
3447 * codepaths and clearing TREE_REF in try_release_extent_buffer.
3448 *
3449 * The actual lifetime of the extent_buffer in the radix tree is
3450 * adequately protected by the refcount, but the TREE_REF bit and
3451 * its corresponding reference are not. To protect against this
3452 * class of races, we call check_buffer_tree_ref from the codepaths
3453 * which trigger io. Note that once io is initiated, TREE_REF can no
3454 * longer be cleared, so that is the moment at which any such race is
3455 * best fixed.
3456 */
3457 refs = atomic_read(&eb->refs);
3458 if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3459 return;
3460
3461 spin_lock(&eb->refs_lock);
3462 if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3463 atomic_inc(&eb->refs);
3464 spin_unlock(&eb->refs_lock);
3465}
3466
3467static void mark_extent_buffer_accessed(struct extent_buffer *eb)
3468{
3469 int num_folios= num_extent_folios(eb);
3470
3471 check_buffer_tree_ref(eb);
3472
3473 for (int i = 0; i < num_folios; i++)
3474 folio_mark_accessed(eb->folios[i]);
3475}
3476
3477struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
3478 u64 start)
3479{
3480 struct extent_buffer *eb;
3481
3482 eb = find_extent_buffer_nolock(fs_info, start);
3483 if (!eb)
3484 return NULL;
3485 /*
3486 * Lock our eb's refs_lock to avoid races with free_extent_buffer().
3487 * When we get our eb it might be flagged with EXTENT_BUFFER_STALE and
3488 * another task running free_extent_buffer() might have seen that flag
3489 * set, eb->refs == 2, that the buffer isn't under IO (dirty and
3490 * writeback flags not set) and it's still in the tree (flag
3491 * EXTENT_BUFFER_TREE_REF set), therefore being in the process of
3492 * decrementing the extent buffer's reference count twice. So here we
3493 * could race and increment the eb's reference count, clear its stale
3494 * flag, mark it as dirty and drop our reference before the other task
3495 * finishes executing free_extent_buffer, which would later result in
3496 * an attempt to free an extent buffer that is dirty.
3497 */
3498 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
3499 spin_lock(&eb->refs_lock);
3500 spin_unlock(&eb->refs_lock);
3501 }
3502 mark_extent_buffer_accessed(eb);
3503 return eb;
3504}
3505
3506#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3507struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
3508 u64 start)
3509{
3510 struct extent_buffer *eb, *exists = NULL;
3511 int ret;
3512
3513 eb = find_extent_buffer(fs_info, start);
3514 if (eb)
3515 return eb;
3516 eb = alloc_dummy_extent_buffer(fs_info, start);
3517 if (!eb)
3518 return ERR_PTR(-ENOMEM);
3519 eb->fs_info = fs_info;
3520again:
3521 ret = radix_tree_preload(GFP_NOFS);
3522 if (ret) {
3523 exists = ERR_PTR(ret);
3524 goto free_eb;
3525 }
3526 spin_lock(&fs_info->buffer_lock);
3527 ret = radix_tree_insert(&fs_info->buffer_radix,
3528 start >> fs_info->sectorsize_bits, eb);
3529 spin_unlock(&fs_info->buffer_lock);
3530 radix_tree_preload_end();
3531 if (ret == -EEXIST) {
3532 exists = find_extent_buffer(fs_info, start);
3533 if (exists)
3534 goto free_eb;
3535 else
3536 goto again;
3537 }
3538 check_buffer_tree_ref(eb);
3539 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
3540
3541 return eb;
3542free_eb:
3543 btrfs_release_extent_buffer(eb);
3544 return exists;
3545}
3546#endif
3547
3548static struct extent_buffer *grab_extent_buffer(
3549 struct btrfs_fs_info *fs_info, struct page *page)
3550{
3551 struct folio *folio = page_folio(page);
3552 struct extent_buffer *exists;
3553
3554 /*
3555 * For subpage case, we completely rely on radix tree to ensure we
3556 * don't try to insert two ebs for the same bytenr. So here we always
3557 * return NULL and just continue.
3558 */
3559 if (fs_info->nodesize < PAGE_SIZE)
3560 return NULL;
3561
3562 /* Page not yet attached to an extent buffer */
3563 if (!folio_test_private(folio))
3564 return NULL;
3565
3566 /*
3567 * We could have already allocated an eb for this page and attached one
3568 * so lets see if we can get a ref on the existing eb, and if we can we
3569 * know it's good and we can just return that one, else we know we can
3570 * just overwrite folio private.
3571 */
3572 exists = folio_get_private(folio);
3573 if (atomic_inc_not_zero(&exists->refs))
3574 return exists;
3575
3576 WARN_ON(PageDirty(page));
3577 folio_detach_private(folio);
3578 return NULL;
3579}
3580
3581static int check_eb_alignment(struct btrfs_fs_info *fs_info, u64 start)
3582{
3583 if (!IS_ALIGNED(start, fs_info->sectorsize)) {
3584 btrfs_err(fs_info, "bad tree block start %llu", start);
3585 return -EINVAL;
3586 }
3587
3588 if (fs_info->nodesize < PAGE_SIZE &&
3589 offset_in_page(start) + fs_info->nodesize > PAGE_SIZE) {
3590 btrfs_err(fs_info,
3591 "tree block crosses page boundary, start %llu nodesize %u",
3592 start, fs_info->nodesize);
3593 return -EINVAL;
3594 }
3595 if (fs_info->nodesize >= PAGE_SIZE &&
3596 !PAGE_ALIGNED(start)) {
3597 btrfs_err(fs_info,
3598 "tree block is not page aligned, start %llu nodesize %u",
3599 start, fs_info->nodesize);
3600 return -EINVAL;
3601 }
3602 if (!IS_ALIGNED(start, fs_info->nodesize) &&
3603 !test_and_set_bit(BTRFS_FS_UNALIGNED_TREE_BLOCK, &fs_info->flags)) {
3604 btrfs_warn(fs_info,
3605"tree block not nodesize aligned, start %llu nodesize %u, can be resolved by a full metadata balance",
3606 start, fs_info->nodesize);
3607 }
3608 return 0;
3609}
3610
3611
3612/*
3613 * Return 0 if eb->folios[i] is attached to btree inode successfully.
3614 * Return >0 if there is already another extent buffer for the range,
3615 * and @found_eb_ret would be updated.
3616 * Return -EAGAIN if the filemap has an existing folio but with different size
3617 * than @eb.
3618 * The caller needs to free the existing folios and retry using the same order.
3619 */
3620static int attach_eb_folio_to_filemap(struct extent_buffer *eb, int i,
3621 struct extent_buffer **found_eb_ret)
3622{
3623
3624 struct btrfs_fs_info *fs_info = eb->fs_info;
3625 struct address_space *mapping = fs_info->btree_inode->i_mapping;
3626 const unsigned long index = eb->start >> PAGE_SHIFT;
3627 struct folio *existing_folio;
3628 int ret;
3629
3630 ASSERT(found_eb_ret);
3631
3632 /* Caller should ensure the folio exists. */
3633 ASSERT(eb->folios[i]);
3634
3635retry:
3636 ret = filemap_add_folio(mapping, eb->folios[i], index + i,
3637 GFP_NOFS | __GFP_NOFAIL);
3638 if (!ret)
3639 return 0;
3640
3641 existing_folio = filemap_lock_folio(mapping, index + i);
3642 /* The page cache only exists for a very short time, just retry. */
3643 if (IS_ERR(existing_folio))
3644 goto retry;
3645
3646 /* For now, we should only have single-page folios for btree inode. */
3647 ASSERT(folio_nr_pages(existing_folio) == 1);
3648
3649 if (folio_size(existing_folio) != folio_size(eb->folios[0])) {
3650 folio_unlock(existing_folio);
3651 folio_put(existing_folio);
3652 return -EAGAIN;
3653 }
3654
3655 if (fs_info->nodesize < PAGE_SIZE) {
3656 /*
3657 * We're going to reuse the existing page, can drop our page
3658 * and subpage structure now.
3659 */
3660 __free_page(folio_page(eb->folios[i], 0));
3661 eb->folios[i] = existing_folio;
3662 } else {
3663 struct extent_buffer *existing_eb;
3664
3665 existing_eb = grab_extent_buffer(fs_info,
3666 folio_page(existing_folio, 0));
3667 if (existing_eb) {
3668 /* The extent buffer still exists, we can use it directly. */
3669 *found_eb_ret = existing_eb;
3670 folio_unlock(existing_folio);
3671 folio_put(existing_folio);
3672 return 1;
3673 }
3674 /* The extent buffer no longer exists, we can reuse the folio. */
3675 __free_page(folio_page(eb->folios[i], 0));
3676 eb->folios[i] = existing_folio;
3677 }
3678 return 0;
3679}
3680
3681struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
3682 u64 start, u64 owner_root, int level)
3683{
3684 unsigned long len = fs_info->nodesize;
3685 int num_folios;
3686 int attached = 0;
3687 struct extent_buffer *eb;
3688 struct extent_buffer *existing_eb = NULL;
3689 struct address_space *mapping = fs_info->btree_inode->i_mapping;
3690 struct btrfs_subpage *prealloc = NULL;
3691 u64 lockdep_owner = owner_root;
3692 bool page_contig = true;
3693 int uptodate = 1;
3694 int ret;
3695
3696 if (check_eb_alignment(fs_info, start))
3697 return ERR_PTR(-EINVAL);
3698
3699#if BITS_PER_LONG == 32
3700 if (start >= MAX_LFS_FILESIZE) {
3701 btrfs_err_rl(fs_info,
3702 "extent buffer %llu is beyond 32bit page cache limit", start);
3703 btrfs_err_32bit_limit(fs_info);
3704 return ERR_PTR(-EOVERFLOW);
3705 }
3706 if (start >= BTRFS_32BIT_EARLY_WARN_THRESHOLD)
3707 btrfs_warn_32bit_limit(fs_info);
3708#endif
3709
3710 eb = find_extent_buffer(fs_info, start);
3711 if (eb)
3712 return eb;
3713
3714 eb = __alloc_extent_buffer(fs_info, start, len);
3715 if (!eb)
3716 return ERR_PTR(-ENOMEM);
3717
3718 /*
3719 * The reloc trees are just snapshots, so we need them to appear to be
3720 * just like any other fs tree WRT lockdep.
3721 */
3722 if (lockdep_owner == BTRFS_TREE_RELOC_OBJECTID)
3723 lockdep_owner = BTRFS_FS_TREE_OBJECTID;
3724
3725 btrfs_set_buffer_lockdep_class(lockdep_owner, eb, level);
3726
3727 /*
3728 * Preallocate folio private for subpage case, so that we won't
3729 * allocate memory with i_private_lock nor page lock hold.
3730 *
3731 * The memory will be freed by attach_extent_buffer_page() or freed
3732 * manually if we exit earlier.
3733 */
3734 if (fs_info->nodesize < PAGE_SIZE) {
3735 prealloc = btrfs_alloc_subpage(fs_info, BTRFS_SUBPAGE_METADATA);
3736 if (IS_ERR(prealloc)) {
3737 ret = PTR_ERR(prealloc);
3738 goto out;
3739 }
3740 }
3741
3742reallocate:
3743 /* Allocate all pages first. */
3744 ret = alloc_eb_folio_array(eb, __GFP_NOFAIL);
3745 if (ret < 0) {
3746 btrfs_free_subpage(prealloc);
3747 goto out;
3748 }
3749
3750 num_folios = num_extent_folios(eb);
3751 /* Attach all pages to the filemap. */
3752 for (int i = 0; i < num_folios; i++) {
3753 struct folio *folio;
3754
3755 ret = attach_eb_folio_to_filemap(eb, i, &existing_eb);
3756 if (ret > 0) {
3757 ASSERT(existing_eb);
3758 goto out;
3759 }
3760
3761 /*
3762 * TODO: Special handling for a corner case where the order of
3763 * folios mismatch between the new eb and filemap.
3764 *
3765 * This happens when:
3766 *
3767 * - the new eb is using higher order folio
3768 *
3769 * - the filemap is still using 0-order folios for the range
3770 * This can happen at the previous eb allocation, and we don't
3771 * have higher order folio for the call.
3772 *
3773 * - the existing eb has already been freed
3774 *
3775 * In this case, we have to free the existing folios first, and
3776 * re-allocate using the same order.
3777 * Thankfully this is not going to happen yet, as we're still
3778 * using 0-order folios.
3779 */
3780 if (unlikely(ret == -EAGAIN)) {
3781 ASSERT(0);
3782 goto reallocate;
3783 }
3784 attached++;
3785
3786 /*
3787 * Only after attach_eb_folio_to_filemap(), eb->folios[] is
3788 * reliable, as we may choose to reuse the existing page cache
3789 * and free the allocated page.
3790 */
3791 folio = eb->folios[i];
3792 spin_lock(&mapping->i_private_lock);
3793 /* Should not fail, as we have preallocated the memory */
3794 ret = attach_extent_buffer_folio(eb, folio, prealloc);
3795 ASSERT(!ret);
3796 /*
3797 * To inform we have extra eb under allocation, so that
3798 * detach_extent_buffer_page() won't release the folio private
3799 * when the eb hasn't yet been inserted into radix tree.
3800 *
3801 * The ref will be decreased when the eb released the page, in
3802 * detach_extent_buffer_page().
3803 * Thus needs no special handling in error path.
3804 */
3805 btrfs_folio_inc_eb_refs(fs_info, folio);
3806 spin_unlock(&mapping->i_private_lock);
3807
3808 WARN_ON(btrfs_folio_test_dirty(fs_info, folio, eb->start, eb->len));
3809
3810 /*
3811 * Check if the current page is physically contiguous with previous eb
3812 * page.
3813 * At this stage, either we allocated a large folio, thus @i
3814 * would only be 0, or we fall back to per-page allocation.
3815 */
3816 if (i && folio_page(eb->folios[i - 1], 0) + 1 != folio_page(folio, 0))
3817 page_contig = false;
3818
3819 if (!btrfs_folio_test_uptodate(fs_info, folio, eb->start, eb->len))
3820 uptodate = 0;
3821
3822 /*
3823 * We can't unlock the pages just yet since the extent buffer
3824 * hasn't been properly inserted in the radix tree, this
3825 * opens a race with btree_release_folio which can free a page
3826 * while we are still filling in all pages for the buffer and
3827 * we could crash.
3828 */
3829 }
3830 if (uptodate)
3831 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3832 /* All pages are physically contiguous, can skip cross page handling. */
3833 if (page_contig)
3834 eb->addr = folio_address(eb->folios[0]) + offset_in_page(eb->start);
3835again:
3836 ret = radix_tree_preload(GFP_NOFS);
3837 if (ret)
3838 goto out;
3839
3840 spin_lock(&fs_info->buffer_lock);
3841 ret = radix_tree_insert(&fs_info->buffer_radix,
3842 start >> fs_info->sectorsize_bits, eb);
3843 spin_unlock(&fs_info->buffer_lock);
3844 radix_tree_preload_end();
3845 if (ret == -EEXIST) {
3846 ret = 0;
3847 existing_eb = find_extent_buffer(fs_info, start);
3848 if (existing_eb)
3849 goto out;
3850 else
3851 goto again;
3852 }
3853 /* add one reference for the tree */
3854 check_buffer_tree_ref(eb);
3855 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
3856
3857 /*
3858 * Now it's safe to unlock the pages because any calls to
3859 * btree_release_folio will correctly detect that a page belongs to a
3860 * live buffer and won't free them prematurely.
3861 */
3862 for (int i = 0; i < num_folios; i++)
3863 unlock_page(folio_page(eb->folios[i], 0));
3864 return eb;
3865
3866out:
3867 WARN_ON(!atomic_dec_and_test(&eb->refs));
3868
3869 /*
3870 * Any attached folios need to be detached before we unlock them. This
3871 * is because when we're inserting our new folios into the mapping, and
3872 * then attaching our eb to that folio. If we fail to insert our folio
3873 * we'll lookup the folio for that index, and grab that EB. We do not
3874 * want that to grab this eb, as we're getting ready to free it. So we
3875 * have to detach it first and then unlock it.
3876 *
3877 * We have to drop our reference and NULL it out here because in the
3878 * subpage case detaching does a btrfs_folio_dec_eb_refs() for our eb.
3879 * Below when we call btrfs_release_extent_buffer() we will call
3880 * detach_extent_buffer_folio() on our remaining pages in the !subpage
3881 * case. If we left eb->folios[i] populated in the subpage case we'd
3882 * double put our reference and be super sad.
3883 */
3884 for (int i = 0; i < attached; i++) {
3885 ASSERT(eb->folios[i]);
3886 detach_extent_buffer_folio(eb, eb->folios[i]);
3887 unlock_page(folio_page(eb->folios[i], 0));
3888 folio_put(eb->folios[i]);
3889 eb->folios[i] = NULL;
3890 }
3891 /*
3892 * Now all pages of that extent buffer is unmapped, set UNMAPPED flag,
3893 * so it can be cleaned up without utlizing page->mapping.
3894 */
3895 set_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
3896
3897 btrfs_release_extent_buffer(eb);
3898 if (ret < 0)
3899 return ERR_PTR(ret);
3900 ASSERT(existing_eb);
3901 return existing_eb;
3902}
3903
3904static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
3905{
3906 struct extent_buffer *eb =
3907 container_of(head, struct extent_buffer, rcu_head);
3908
3909 __free_extent_buffer(eb);
3910}
3911
3912static int release_extent_buffer(struct extent_buffer *eb)
3913 __releases(&eb->refs_lock)
3914{
3915 lockdep_assert_held(&eb->refs_lock);
3916
3917 WARN_ON(atomic_read(&eb->refs) == 0);
3918 if (atomic_dec_and_test(&eb->refs)) {
3919 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
3920 struct btrfs_fs_info *fs_info = eb->fs_info;
3921
3922 spin_unlock(&eb->refs_lock);
3923
3924 spin_lock(&fs_info->buffer_lock);
3925 radix_tree_delete(&fs_info->buffer_radix,
3926 eb->start >> fs_info->sectorsize_bits);
3927 spin_unlock(&fs_info->buffer_lock);
3928 } else {
3929 spin_unlock(&eb->refs_lock);
3930 }
3931
3932 btrfs_leak_debug_del_eb(eb);
3933 /* Should be safe to release our pages at this point */
3934 btrfs_release_extent_buffer_pages(eb);
3935#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3936 if (unlikely(test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags))) {
3937 __free_extent_buffer(eb);
3938 return 1;
3939 }
3940#endif
3941 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
3942 return 1;
3943 }
3944 spin_unlock(&eb->refs_lock);
3945
3946 return 0;
3947}
3948
3949void free_extent_buffer(struct extent_buffer *eb)
3950{
3951 int refs;
3952 if (!eb)
3953 return;
3954
3955 refs = atomic_read(&eb->refs);
3956 while (1) {
3957 if ((!test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) && refs <= 3)
3958 || (test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) &&
3959 refs == 1))
3960 break;
3961 if (atomic_try_cmpxchg(&eb->refs, &refs, refs - 1))
3962 return;
3963 }
3964
3965 spin_lock(&eb->refs_lock);
3966 if (atomic_read(&eb->refs) == 2 &&
3967 test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
3968 !extent_buffer_under_io(eb) &&
3969 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3970 atomic_dec(&eb->refs);
3971
3972 /*
3973 * I know this is terrible, but it's temporary until we stop tracking
3974 * the uptodate bits and such for the extent buffers.
3975 */
3976 release_extent_buffer(eb);
3977}
3978
3979void free_extent_buffer_stale(struct extent_buffer *eb)
3980{
3981 if (!eb)
3982 return;
3983
3984 spin_lock(&eb->refs_lock);
3985 set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
3986
3987 if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
3988 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3989 atomic_dec(&eb->refs);
3990 release_extent_buffer(eb);
3991}
3992
3993static void btree_clear_folio_dirty(struct folio *folio)
3994{
3995 ASSERT(folio_test_dirty(folio));
3996 ASSERT(folio_test_locked(folio));
3997 folio_clear_dirty_for_io(folio);
3998 xa_lock_irq(&folio->mapping->i_pages);
3999 if (!folio_test_dirty(folio))
4000 __xa_clear_mark(&folio->mapping->i_pages,
4001 folio_index(folio), PAGECACHE_TAG_DIRTY);
4002 xa_unlock_irq(&folio->mapping->i_pages);
4003}
4004
4005static void clear_subpage_extent_buffer_dirty(const struct extent_buffer *eb)
4006{
4007 struct btrfs_fs_info *fs_info = eb->fs_info;
4008 struct folio *folio = eb->folios[0];
4009 bool last;
4010
4011 /* btree_clear_folio_dirty() needs page locked. */
4012 folio_lock(folio);
4013 last = btrfs_subpage_clear_and_test_dirty(fs_info, folio, eb->start, eb->len);
4014 if (last)
4015 btree_clear_folio_dirty(folio);
4016 folio_unlock(folio);
4017 WARN_ON(atomic_read(&eb->refs) == 0);
4018}
4019
4020void btrfs_clear_buffer_dirty(struct btrfs_trans_handle *trans,
4021 struct extent_buffer *eb)
4022{
4023 struct btrfs_fs_info *fs_info = eb->fs_info;
4024 int num_folios;
4025
4026 btrfs_assert_tree_write_locked(eb);
4027
4028 if (trans && btrfs_header_generation(eb) != trans->transid)
4029 return;
4030
4031 /*
4032 * Instead of clearing the dirty flag off of the buffer, mark it as
4033 * EXTENT_BUFFER_ZONED_ZEROOUT. This allows us to preserve
4034 * write-ordering in zoned mode, without the need to later re-dirty
4035 * the extent_buffer.
4036 *
4037 * The actual zeroout of the buffer will happen later in
4038 * btree_csum_one_bio.
4039 */
4040 if (btrfs_is_zoned(fs_info)) {
4041 set_bit(EXTENT_BUFFER_ZONED_ZEROOUT, &eb->bflags);
4042 return;
4043 }
4044
4045 if (!test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags))
4046 return;
4047
4048 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes, -eb->len,
4049 fs_info->dirty_metadata_batch);
4050
4051 if (eb->fs_info->nodesize < PAGE_SIZE)
4052 return clear_subpage_extent_buffer_dirty(eb);
4053
4054 num_folios = num_extent_folios(eb);
4055 for (int i = 0; i < num_folios; i++) {
4056 struct folio *folio = eb->folios[i];
4057
4058 if (!folio_test_dirty(folio))
4059 continue;
4060 folio_lock(folio);
4061 btree_clear_folio_dirty(folio);
4062 folio_unlock(folio);
4063 }
4064 WARN_ON(atomic_read(&eb->refs) == 0);
4065}
4066
4067void set_extent_buffer_dirty(struct extent_buffer *eb)
4068{
4069 int num_folios;
4070 bool was_dirty;
4071
4072 check_buffer_tree_ref(eb);
4073
4074 was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4075
4076 num_folios = num_extent_folios(eb);
4077 WARN_ON(atomic_read(&eb->refs) == 0);
4078 WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
4079
4080 if (!was_dirty) {
4081 bool subpage = eb->fs_info->nodesize < PAGE_SIZE;
4082
4083 /*
4084 * For subpage case, we can have other extent buffers in the
4085 * same page, and in clear_subpage_extent_buffer_dirty() we
4086 * have to clear page dirty without subpage lock held.
4087 * This can cause race where our page gets dirty cleared after
4088 * we just set it.
4089 *
4090 * Thankfully, clear_subpage_extent_buffer_dirty() has locked
4091 * its page for other reasons, we can use page lock to prevent
4092 * the above race.
4093 */
4094 if (subpage)
4095 lock_page(folio_page(eb->folios[0], 0));
4096 for (int i = 0; i < num_folios; i++)
4097 btrfs_folio_set_dirty(eb->fs_info, eb->folios[i],
4098 eb->start, eb->len);
4099 if (subpage)
4100 unlock_page(folio_page(eb->folios[0], 0));
4101 percpu_counter_add_batch(&eb->fs_info->dirty_metadata_bytes,
4102 eb->len,
4103 eb->fs_info->dirty_metadata_batch);
4104 }
4105#ifdef CONFIG_BTRFS_DEBUG
4106 for (int i = 0; i < num_folios; i++)
4107 ASSERT(folio_test_dirty(eb->folios[i]));
4108#endif
4109}
4110
4111void clear_extent_buffer_uptodate(struct extent_buffer *eb)
4112{
4113 struct btrfs_fs_info *fs_info = eb->fs_info;
4114 int num_folios = num_extent_folios(eb);
4115
4116 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4117 for (int i = 0; i < num_folios; i++) {
4118 struct folio *folio = eb->folios[i];
4119
4120 if (!folio)
4121 continue;
4122
4123 /*
4124 * This is special handling for metadata subpage, as regular
4125 * btrfs_is_subpage() can not handle cloned/dummy metadata.
4126 */
4127 if (fs_info->nodesize >= PAGE_SIZE)
4128 folio_clear_uptodate(folio);
4129 else
4130 btrfs_subpage_clear_uptodate(fs_info, folio,
4131 eb->start, eb->len);
4132 }
4133}
4134
4135void set_extent_buffer_uptodate(struct extent_buffer *eb)
4136{
4137 struct btrfs_fs_info *fs_info = eb->fs_info;
4138 int num_folios = num_extent_folios(eb);
4139
4140 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4141 for (int i = 0; i < num_folios; i++) {
4142 struct folio *folio = eb->folios[i];
4143
4144 /*
4145 * This is special handling for metadata subpage, as regular
4146 * btrfs_is_subpage() can not handle cloned/dummy metadata.
4147 */
4148 if (fs_info->nodesize >= PAGE_SIZE)
4149 folio_mark_uptodate(folio);
4150 else
4151 btrfs_subpage_set_uptodate(fs_info, folio,
4152 eb->start, eb->len);
4153 }
4154}
4155
4156static void end_bbio_meta_read(struct btrfs_bio *bbio)
4157{
4158 struct extent_buffer *eb = bbio->private;
4159 struct btrfs_fs_info *fs_info = eb->fs_info;
4160 bool uptodate = !bbio->bio.bi_status;
4161 struct folio_iter fi;
4162 u32 bio_offset = 0;
4163
4164 eb->read_mirror = bbio->mirror_num;
4165
4166 if (uptodate &&
4167 btrfs_validate_extent_buffer(eb, &bbio->parent_check) < 0)
4168 uptodate = false;
4169
4170 if (uptodate) {
4171 set_extent_buffer_uptodate(eb);
4172 } else {
4173 clear_extent_buffer_uptodate(eb);
4174 set_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
4175 }
4176
4177 bio_for_each_folio_all(fi, &bbio->bio) {
4178 struct folio *folio = fi.folio;
4179 u64 start = eb->start + bio_offset;
4180 u32 len = fi.length;
4181
4182 if (uptodate)
4183 btrfs_folio_set_uptodate(fs_info, folio, start, len);
4184 else
4185 btrfs_folio_clear_uptodate(fs_info, folio, start, len);
4186
4187 bio_offset += len;
4188 }
4189
4190 clear_bit(EXTENT_BUFFER_READING, &eb->bflags);
4191 smp_mb__after_atomic();
4192 wake_up_bit(&eb->bflags, EXTENT_BUFFER_READING);
4193 free_extent_buffer(eb);
4194
4195 bio_put(&bbio->bio);
4196}
4197
4198int read_extent_buffer_pages(struct extent_buffer *eb, int wait, int mirror_num,
4199 struct btrfs_tree_parent_check *check)
4200{
4201 struct btrfs_bio *bbio;
4202 bool ret;
4203
4204 if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4205 return 0;
4206
4207 /*
4208 * We could have had EXTENT_BUFFER_UPTODATE cleared by the write
4209 * operation, which could potentially still be in flight. In this case
4210 * we simply want to return an error.
4211 */
4212 if (unlikely(test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)))
4213 return -EIO;
4214
4215 /* Someone else is already reading the buffer, just wait for it. */
4216 if (test_and_set_bit(EXTENT_BUFFER_READING, &eb->bflags))
4217 goto done;
4218
4219 clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
4220 eb->read_mirror = 0;
4221 check_buffer_tree_ref(eb);
4222 atomic_inc(&eb->refs);
4223
4224 bbio = btrfs_bio_alloc(INLINE_EXTENT_BUFFER_PAGES,
4225 REQ_OP_READ | REQ_META, eb->fs_info,
4226 end_bbio_meta_read, eb);
4227 bbio->bio.bi_iter.bi_sector = eb->start >> SECTOR_SHIFT;
4228 bbio->inode = BTRFS_I(eb->fs_info->btree_inode);
4229 bbio->file_offset = eb->start;
4230 memcpy(&bbio->parent_check, check, sizeof(*check));
4231 if (eb->fs_info->nodesize < PAGE_SIZE) {
4232 ret = bio_add_folio(&bbio->bio, eb->folios[0], eb->len,
4233 eb->start - folio_pos(eb->folios[0]));
4234 ASSERT(ret);
4235 } else {
4236 int num_folios = num_extent_folios(eb);
4237
4238 for (int i = 0; i < num_folios; i++) {
4239 struct folio *folio = eb->folios[i];
4240
4241 ret = bio_add_folio(&bbio->bio, folio, folio_size(folio), 0);
4242 ASSERT(ret);
4243 }
4244 }
4245 btrfs_submit_bio(bbio, mirror_num);
4246
4247done:
4248 if (wait == WAIT_COMPLETE) {
4249 wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_READING, TASK_UNINTERRUPTIBLE);
4250 if (!test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4251 return -EIO;
4252 }
4253
4254 return 0;
4255}
4256
4257static bool report_eb_range(const struct extent_buffer *eb, unsigned long start,
4258 unsigned long len)
4259{
4260 btrfs_warn(eb->fs_info,
4261 "access to eb bytenr %llu len %lu out of range start %lu len %lu",
4262 eb->start, eb->len, start, len);
4263 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
4264
4265 return true;
4266}
4267
4268/*
4269 * Check if the [start, start + len) range is valid before reading/writing
4270 * the eb.
4271 * NOTE: @start and @len are offset inside the eb, not logical address.
4272 *
4273 * Caller should not touch the dst/src memory if this function returns error.
4274 */
4275static inline int check_eb_range(const struct extent_buffer *eb,
4276 unsigned long start, unsigned long len)
4277{
4278 unsigned long offset;
4279
4280 /* start, start + len should not go beyond eb->len nor overflow */
4281 if (unlikely(check_add_overflow(start, len, &offset) || offset > eb->len))
4282 return report_eb_range(eb, start, len);
4283
4284 return false;
4285}
4286
4287void read_extent_buffer(const struct extent_buffer *eb, void *dstv,
4288 unsigned long start, unsigned long len)
4289{
4290 const int unit_size = folio_size(eb->folios[0]);
4291 size_t cur;
4292 size_t offset;
4293 char *dst = (char *)dstv;
4294 unsigned long i = get_eb_folio_index(eb, start);
4295
4296 if (check_eb_range(eb, start, len)) {
4297 /*
4298 * Invalid range hit, reset the memory, so callers won't get
4299 * some random garbage for their uninitialized memory.
4300 */
4301 memset(dstv, 0, len);
4302 return;
4303 }
4304
4305 if (eb->addr) {
4306 memcpy(dstv, eb->addr + start, len);
4307 return;
4308 }
4309
4310 offset = get_eb_offset_in_folio(eb, start);
4311
4312 while (len > 0) {
4313 char *kaddr;
4314
4315 cur = min(len, unit_size - offset);
4316 kaddr = folio_address(eb->folios[i]);
4317 memcpy(dst, kaddr + offset, cur);
4318
4319 dst += cur;
4320 len -= cur;
4321 offset = 0;
4322 i++;
4323 }
4324}
4325
4326int read_extent_buffer_to_user_nofault(const struct extent_buffer *eb,
4327 void __user *dstv,
4328 unsigned long start, unsigned long len)
4329{
4330 const int unit_size = folio_size(eb->folios[0]);
4331 size_t cur;
4332 size_t offset;
4333 char __user *dst = (char __user *)dstv;
4334 unsigned long i = get_eb_folio_index(eb, start);
4335 int ret = 0;
4336
4337 WARN_ON(start > eb->len);
4338 WARN_ON(start + len > eb->start + eb->len);
4339
4340 if (eb->addr) {
4341 if (copy_to_user_nofault(dstv, eb->addr + start, len))
4342 ret = -EFAULT;
4343 return ret;
4344 }
4345
4346 offset = get_eb_offset_in_folio(eb, start);
4347
4348 while (len > 0) {
4349 char *kaddr;
4350
4351 cur = min(len, unit_size - offset);
4352 kaddr = folio_address(eb->folios[i]);
4353 if (copy_to_user_nofault(dst, kaddr + offset, cur)) {
4354 ret = -EFAULT;
4355 break;
4356 }
4357
4358 dst += cur;
4359 len -= cur;
4360 offset = 0;
4361 i++;
4362 }
4363
4364 return ret;
4365}
4366
4367int memcmp_extent_buffer(const struct extent_buffer *eb, const void *ptrv,
4368 unsigned long start, unsigned long len)
4369{
4370 const int unit_size = folio_size(eb->folios[0]);
4371 size_t cur;
4372 size_t offset;
4373 char *kaddr;
4374 char *ptr = (char *)ptrv;
4375 unsigned long i = get_eb_folio_index(eb, start);
4376 int ret = 0;
4377
4378 if (check_eb_range(eb, start, len))
4379 return -EINVAL;
4380
4381 if (eb->addr)
4382 return memcmp(ptrv, eb->addr + start, len);
4383
4384 offset = get_eb_offset_in_folio(eb, start);
4385
4386 while (len > 0) {
4387 cur = min(len, unit_size - offset);
4388 kaddr = folio_address(eb->folios[i]);
4389 ret = memcmp(ptr, kaddr + offset, cur);
4390 if (ret)
4391 break;
4392
4393 ptr += cur;
4394 len -= cur;
4395 offset = 0;
4396 i++;
4397 }
4398 return ret;
4399}
4400
4401/*
4402 * Check that the extent buffer is uptodate.
4403 *
4404 * For regular sector size == PAGE_SIZE case, check if @page is uptodate.
4405 * For subpage case, check if the range covered by the eb has EXTENT_UPTODATE.
4406 */
4407static void assert_eb_folio_uptodate(const struct extent_buffer *eb, int i)
4408{
4409 struct btrfs_fs_info *fs_info = eb->fs_info;
4410 struct folio *folio = eb->folios[i];
4411
4412 ASSERT(folio);
4413
4414 /*
4415 * If we are using the commit root we could potentially clear a page
4416 * Uptodate while we're using the extent buffer that we've previously
4417 * looked up. We don't want to complain in this case, as the page was
4418 * valid before, we just didn't write it out. Instead we want to catch
4419 * the case where we didn't actually read the block properly, which
4420 * would have !PageUptodate and !EXTENT_BUFFER_WRITE_ERR.
4421 */
4422 if (test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
4423 return;
4424
4425 if (fs_info->nodesize < PAGE_SIZE) {
4426 struct folio *folio = eb->folios[0];
4427
4428 ASSERT(i == 0);
4429 if (WARN_ON(!btrfs_subpage_test_uptodate(fs_info, folio,
4430 eb->start, eb->len)))
4431 btrfs_subpage_dump_bitmap(fs_info, folio, eb->start, eb->len);
4432 } else {
4433 WARN_ON(!folio_test_uptodate(folio));
4434 }
4435}
4436
4437static void __write_extent_buffer(const struct extent_buffer *eb,
4438 const void *srcv, unsigned long start,
4439 unsigned long len, bool use_memmove)
4440{
4441 const int unit_size = folio_size(eb->folios[0]);
4442 size_t cur;
4443 size_t offset;
4444 char *kaddr;
4445 char *src = (char *)srcv;
4446 unsigned long i = get_eb_folio_index(eb, start);
4447 /* For unmapped (dummy) ebs, no need to check their uptodate status. */
4448 const bool check_uptodate = !test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
4449
4450 if (check_eb_range(eb, start, len))
4451 return;
4452
4453 if (eb->addr) {
4454 if (use_memmove)
4455 memmove(eb->addr + start, srcv, len);
4456 else
4457 memcpy(eb->addr + start, srcv, len);
4458 return;
4459 }
4460
4461 offset = get_eb_offset_in_folio(eb, start);
4462
4463 while (len > 0) {
4464 if (check_uptodate)
4465 assert_eb_folio_uptodate(eb, i);
4466
4467 cur = min(len, unit_size - offset);
4468 kaddr = folio_address(eb->folios[i]);
4469 if (use_memmove)
4470 memmove(kaddr + offset, src, cur);
4471 else
4472 memcpy(kaddr + offset, src, cur);
4473
4474 src += cur;
4475 len -= cur;
4476 offset = 0;
4477 i++;
4478 }
4479}
4480
4481void write_extent_buffer(const struct extent_buffer *eb, const void *srcv,
4482 unsigned long start, unsigned long len)
4483{
4484 return __write_extent_buffer(eb, srcv, start, len, false);
4485}
4486
4487static void memset_extent_buffer(const struct extent_buffer *eb, int c,
4488 unsigned long start, unsigned long len)
4489{
4490 const int unit_size = folio_size(eb->folios[0]);
4491 unsigned long cur = start;
4492
4493 if (eb->addr) {
4494 memset(eb->addr + start, c, len);
4495 return;
4496 }
4497
4498 while (cur < start + len) {
4499 unsigned long index = get_eb_folio_index(eb, cur);
4500 unsigned int offset = get_eb_offset_in_folio(eb, cur);
4501 unsigned int cur_len = min(start + len - cur, unit_size - offset);
4502
4503 assert_eb_folio_uptodate(eb, index);
4504 memset(folio_address(eb->folios[index]) + offset, c, cur_len);
4505
4506 cur += cur_len;
4507 }
4508}
4509
4510void memzero_extent_buffer(const struct extent_buffer *eb, unsigned long start,
4511 unsigned long len)
4512{
4513 if (check_eb_range(eb, start, len))
4514 return;
4515 return memset_extent_buffer(eb, 0, start, len);
4516}
4517
4518void copy_extent_buffer_full(const struct extent_buffer *dst,
4519 const struct extent_buffer *src)
4520{
4521 const int unit_size = folio_size(src->folios[0]);
4522 unsigned long cur = 0;
4523
4524 ASSERT(dst->len == src->len);
4525
4526 while (cur < src->len) {
4527 unsigned long index = get_eb_folio_index(src, cur);
4528 unsigned long offset = get_eb_offset_in_folio(src, cur);
4529 unsigned long cur_len = min(src->len, unit_size - offset);
4530 void *addr = folio_address(src->folios[index]) + offset;
4531
4532 write_extent_buffer(dst, addr, cur, cur_len);
4533
4534 cur += cur_len;
4535 }
4536}
4537
4538void copy_extent_buffer(const struct extent_buffer *dst,
4539 const struct extent_buffer *src,
4540 unsigned long dst_offset, unsigned long src_offset,
4541 unsigned long len)
4542{
4543 const int unit_size = folio_size(dst->folios[0]);
4544 u64 dst_len = dst->len;
4545 size_t cur;
4546 size_t offset;
4547 char *kaddr;
4548 unsigned long i = get_eb_folio_index(dst, dst_offset);
4549
4550 if (check_eb_range(dst, dst_offset, len) ||
4551 check_eb_range(src, src_offset, len))
4552 return;
4553
4554 WARN_ON(src->len != dst_len);
4555
4556 offset = get_eb_offset_in_folio(dst, dst_offset);
4557
4558 while (len > 0) {
4559 assert_eb_folio_uptodate(dst, i);
4560
4561 cur = min(len, (unsigned long)(unit_size - offset));
4562
4563 kaddr = folio_address(dst->folios[i]);
4564 read_extent_buffer(src, kaddr + offset, src_offset, cur);
4565
4566 src_offset += cur;
4567 len -= cur;
4568 offset = 0;
4569 i++;
4570 }
4571}
4572
4573/*
4574 * Calculate the folio and offset of the byte containing the given bit number.
4575 *
4576 * @eb: the extent buffer
4577 * @start: offset of the bitmap item in the extent buffer
4578 * @nr: bit number
4579 * @folio_index: return index of the folio in the extent buffer that contains
4580 * the given bit number
4581 * @folio_offset: return offset into the folio given by folio_index
4582 *
4583 * This helper hides the ugliness of finding the byte in an extent buffer which
4584 * contains a given bit.
4585 */
4586static inline void eb_bitmap_offset(const struct extent_buffer *eb,
4587 unsigned long start, unsigned long nr,
4588 unsigned long *folio_index,
4589 size_t *folio_offset)
4590{
4591 size_t byte_offset = BIT_BYTE(nr);
4592 size_t offset;
4593
4594 /*
4595 * The byte we want is the offset of the extent buffer + the offset of
4596 * the bitmap item in the extent buffer + the offset of the byte in the
4597 * bitmap item.
4598 */
4599 offset = start + offset_in_folio(eb->folios[0], eb->start) + byte_offset;
4600
4601 *folio_index = offset >> folio_shift(eb->folios[0]);
4602 *folio_offset = offset_in_folio(eb->folios[0], offset);
4603}
4604
4605/*
4606 * Determine whether a bit in a bitmap item is set.
4607 *
4608 * @eb: the extent buffer
4609 * @start: offset of the bitmap item in the extent buffer
4610 * @nr: bit number to test
4611 */
4612int extent_buffer_test_bit(const struct extent_buffer *eb, unsigned long start,
4613 unsigned long nr)
4614{
4615 unsigned long i;
4616 size_t offset;
4617 u8 *kaddr;
4618
4619 eb_bitmap_offset(eb, start, nr, &i, &offset);
4620 assert_eb_folio_uptodate(eb, i);
4621 kaddr = folio_address(eb->folios[i]);
4622 return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
4623}
4624
4625static u8 *extent_buffer_get_byte(const struct extent_buffer *eb, unsigned long bytenr)
4626{
4627 unsigned long index = get_eb_folio_index(eb, bytenr);
4628
4629 if (check_eb_range(eb, bytenr, 1))
4630 return NULL;
4631 return folio_address(eb->folios[index]) + get_eb_offset_in_folio(eb, bytenr);
4632}
4633
4634/*
4635 * Set an area of a bitmap to 1.
4636 *
4637 * @eb: the extent buffer
4638 * @start: offset of the bitmap item in the extent buffer
4639 * @pos: bit number of the first bit
4640 * @len: number of bits to set
4641 */
4642void extent_buffer_bitmap_set(const struct extent_buffer *eb, unsigned long start,
4643 unsigned long pos, unsigned long len)
4644{
4645 unsigned int first_byte = start + BIT_BYTE(pos);
4646 unsigned int last_byte = start + BIT_BYTE(pos + len - 1);
4647 const bool same_byte = (first_byte == last_byte);
4648 u8 mask = BITMAP_FIRST_BYTE_MASK(pos);
4649 u8 *kaddr;
4650
4651 if (same_byte)
4652 mask &= BITMAP_LAST_BYTE_MASK(pos + len);
4653
4654 /* Handle the first byte. */
4655 kaddr = extent_buffer_get_byte(eb, first_byte);
4656 *kaddr |= mask;
4657 if (same_byte)
4658 return;
4659
4660 /* Handle the byte aligned part. */
4661 ASSERT(first_byte + 1 <= last_byte);
4662 memset_extent_buffer(eb, 0xff, first_byte + 1, last_byte - first_byte - 1);
4663
4664 /* Handle the last byte. */
4665 kaddr = extent_buffer_get_byte(eb, last_byte);
4666 *kaddr |= BITMAP_LAST_BYTE_MASK(pos + len);
4667}
4668
4669
4670/*
4671 * Clear an area of a bitmap.
4672 *
4673 * @eb: the extent buffer
4674 * @start: offset of the bitmap item in the extent buffer
4675 * @pos: bit number of the first bit
4676 * @len: number of bits to clear
4677 */
4678void extent_buffer_bitmap_clear(const struct extent_buffer *eb,
4679 unsigned long start, unsigned long pos,
4680 unsigned long len)
4681{
4682 unsigned int first_byte = start + BIT_BYTE(pos);
4683 unsigned int last_byte = start + BIT_BYTE(pos + len - 1);
4684 const bool same_byte = (first_byte == last_byte);
4685 u8 mask = BITMAP_FIRST_BYTE_MASK(pos);
4686 u8 *kaddr;
4687
4688 if (same_byte)
4689 mask &= BITMAP_LAST_BYTE_MASK(pos + len);
4690
4691 /* Handle the first byte. */
4692 kaddr = extent_buffer_get_byte(eb, first_byte);
4693 *kaddr &= ~mask;
4694 if (same_byte)
4695 return;
4696
4697 /* Handle the byte aligned part. */
4698 ASSERT(first_byte + 1 <= last_byte);
4699 memset_extent_buffer(eb, 0, first_byte + 1, last_byte - first_byte - 1);
4700
4701 /* Handle the last byte. */
4702 kaddr = extent_buffer_get_byte(eb, last_byte);
4703 *kaddr &= ~BITMAP_LAST_BYTE_MASK(pos + len);
4704}
4705
4706static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
4707{
4708 unsigned long distance = (src > dst) ? src - dst : dst - src;
4709 return distance < len;
4710}
4711
4712void memcpy_extent_buffer(const struct extent_buffer *dst,
4713 unsigned long dst_offset, unsigned long src_offset,
4714 unsigned long len)
4715{
4716 const int unit_size = folio_size(dst->folios[0]);
4717 unsigned long cur_off = 0;
4718
4719 if (check_eb_range(dst, dst_offset, len) ||
4720 check_eb_range(dst, src_offset, len))
4721 return;
4722
4723 if (dst->addr) {
4724 const bool use_memmove = areas_overlap(src_offset, dst_offset, len);
4725
4726 if (use_memmove)
4727 memmove(dst->addr + dst_offset, dst->addr + src_offset, len);
4728 else
4729 memcpy(dst->addr + dst_offset, dst->addr + src_offset, len);
4730 return;
4731 }
4732
4733 while (cur_off < len) {
4734 unsigned long cur_src = cur_off + src_offset;
4735 unsigned long folio_index = get_eb_folio_index(dst, cur_src);
4736 unsigned long folio_off = get_eb_offset_in_folio(dst, cur_src);
4737 unsigned long cur_len = min(src_offset + len - cur_src,
4738 unit_size - folio_off);
4739 void *src_addr = folio_address(dst->folios[folio_index]) + folio_off;
4740 const bool use_memmove = areas_overlap(src_offset + cur_off,
4741 dst_offset + cur_off, cur_len);
4742
4743 __write_extent_buffer(dst, src_addr, dst_offset + cur_off, cur_len,
4744 use_memmove);
4745 cur_off += cur_len;
4746 }
4747}
4748
4749void memmove_extent_buffer(const struct extent_buffer *dst,
4750 unsigned long dst_offset, unsigned long src_offset,
4751 unsigned long len)
4752{
4753 unsigned long dst_end = dst_offset + len - 1;
4754 unsigned long src_end = src_offset + len - 1;
4755
4756 if (check_eb_range(dst, dst_offset, len) ||
4757 check_eb_range(dst, src_offset, len))
4758 return;
4759
4760 if (dst_offset < src_offset) {
4761 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
4762 return;
4763 }
4764
4765 if (dst->addr) {
4766 memmove(dst->addr + dst_offset, dst->addr + src_offset, len);
4767 return;
4768 }
4769
4770 while (len > 0) {
4771 unsigned long src_i;
4772 size_t cur;
4773 size_t dst_off_in_folio;
4774 size_t src_off_in_folio;
4775 void *src_addr;
4776 bool use_memmove;
4777
4778 src_i = get_eb_folio_index(dst, src_end);
4779
4780 dst_off_in_folio = get_eb_offset_in_folio(dst, dst_end);
4781 src_off_in_folio = get_eb_offset_in_folio(dst, src_end);
4782
4783 cur = min_t(unsigned long, len, src_off_in_folio + 1);
4784 cur = min(cur, dst_off_in_folio + 1);
4785
4786 src_addr = folio_address(dst->folios[src_i]) + src_off_in_folio -
4787 cur + 1;
4788 use_memmove = areas_overlap(src_end - cur + 1, dst_end - cur + 1,
4789 cur);
4790
4791 __write_extent_buffer(dst, src_addr, dst_end - cur + 1, cur,
4792 use_memmove);
4793
4794 dst_end -= cur;
4795 src_end -= cur;
4796 len -= cur;
4797 }
4798}
4799
4800#define GANG_LOOKUP_SIZE 16
4801static struct extent_buffer *get_next_extent_buffer(
4802 struct btrfs_fs_info *fs_info, struct page *page, u64 bytenr)
4803{
4804 struct extent_buffer *gang[GANG_LOOKUP_SIZE];
4805 struct extent_buffer *found = NULL;
4806 u64 page_start = page_offset(page);
4807 u64 cur = page_start;
4808
4809 ASSERT(in_range(bytenr, page_start, PAGE_SIZE));
4810 lockdep_assert_held(&fs_info->buffer_lock);
4811
4812 while (cur < page_start + PAGE_SIZE) {
4813 int ret;
4814 int i;
4815
4816 ret = radix_tree_gang_lookup(&fs_info->buffer_radix,
4817 (void **)gang, cur >> fs_info->sectorsize_bits,
4818 min_t(unsigned int, GANG_LOOKUP_SIZE,
4819 PAGE_SIZE / fs_info->nodesize));
4820 if (ret == 0)
4821 goto out;
4822 for (i = 0; i < ret; i++) {
4823 /* Already beyond page end */
4824 if (gang[i]->start >= page_start + PAGE_SIZE)
4825 goto out;
4826 /* Found one */
4827 if (gang[i]->start >= bytenr) {
4828 found = gang[i];
4829 goto out;
4830 }
4831 }
4832 cur = gang[ret - 1]->start + gang[ret - 1]->len;
4833 }
4834out:
4835 return found;
4836}
4837
4838static int try_release_subpage_extent_buffer(struct page *page)
4839{
4840 struct btrfs_fs_info *fs_info = btrfs_sb(page->mapping->host->i_sb);
4841 u64 cur = page_offset(page);
4842 const u64 end = page_offset(page) + PAGE_SIZE;
4843 int ret;
4844
4845 while (cur < end) {
4846 struct extent_buffer *eb = NULL;
4847
4848 /*
4849 * Unlike try_release_extent_buffer() which uses folio private
4850 * to grab buffer, for subpage case we rely on radix tree, thus
4851 * we need to ensure radix tree consistency.
4852 *
4853 * We also want an atomic snapshot of the radix tree, thus go
4854 * with spinlock rather than RCU.
4855 */
4856 spin_lock(&fs_info->buffer_lock);
4857 eb = get_next_extent_buffer(fs_info, page, cur);
4858 if (!eb) {
4859 /* No more eb in the page range after or at cur */
4860 spin_unlock(&fs_info->buffer_lock);
4861 break;
4862 }
4863 cur = eb->start + eb->len;
4864
4865 /*
4866 * The same as try_release_extent_buffer(), to ensure the eb
4867 * won't disappear out from under us.
4868 */
4869 spin_lock(&eb->refs_lock);
4870 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
4871 spin_unlock(&eb->refs_lock);
4872 spin_unlock(&fs_info->buffer_lock);
4873 break;
4874 }
4875 spin_unlock(&fs_info->buffer_lock);
4876
4877 /*
4878 * If tree ref isn't set then we know the ref on this eb is a
4879 * real ref, so just return, this eb will likely be freed soon
4880 * anyway.
4881 */
4882 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4883 spin_unlock(&eb->refs_lock);
4884 break;
4885 }
4886
4887 /*
4888 * Here we don't care about the return value, we will always
4889 * check the folio private at the end. And
4890 * release_extent_buffer() will release the refs_lock.
4891 */
4892 release_extent_buffer(eb);
4893 }
4894 /*
4895 * Finally to check if we have cleared folio private, as if we have
4896 * released all ebs in the page, the folio private should be cleared now.
4897 */
4898 spin_lock(&page->mapping->i_private_lock);
4899 if (!folio_test_private(page_folio(page)))
4900 ret = 1;
4901 else
4902 ret = 0;
4903 spin_unlock(&page->mapping->i_private_lock);
4904 return ret;
4905
4906}
4907
4908int try_release_extent_buffer(struct page *page)
4909{
4910 struct folio *folio = page_folio(page);
4911 struct extent_buffer *eb;
4912
4913 if (btrfs_sb(page->mapping->host->i_sb)->nodesize < PAGE_SIZE)
4914 return try_release_subpage_extent_buffer(page);
4915
4916 /*
4917 * We need to make sure nobody is changing folio private, as we rely on
4918 * folio private as the pointer to extent buffer.
4919 */
4920 spin_lock(&page->mapping->i_private_lock);
4921 if (!folio_test_private(folio)) {
4922 spin_unlock(&page->mapping->i_private_lock);
4923 return 1;
4924 }
4925
4926 eb = folio_get_private(folio);
4927 BUG_ON(!eb);
4928
4929 /*
4930 * This is a little awful but should be ok, we need to make sure that
4931 * the eb doesn't disappear out from under us while we're looking at
4932 * this page.
4933 */
4934 spin_lock(&eb->refs_lock);
4935 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
4936 spin_unlock(&eb->refs_lock);
4937 spin_unlock(&page->mapping->i_private_lock);
4938 return 0;
4939 }
4940 spin_unlock(&page->mapping->i_private_lock);
4941
4942 /*
4943 * If tree ref isn't set then we know the ref on this eb is a real ref,
4944 * so just return, this page will likely be freed soon anyway.
4945 */
4946 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4947 spin_unlock(&eb->refs_lock);
4948 return 0;
4949 }
4950
4951 return release_extent_buffer(eb);
4952}
4953
4954/*
4955 * Attempt to readahead a child block.
4956 *
4957 * @fs_info: the fs_info
4958 * @bytenr: bytenr to read
4959 * @owner_root: objectid of the root that owns this eb
4960 * @gen: generation for the uptodate check, can be 0
4961 * @level: level for the eb
4962 *
4963 * Attempt to readahead a tree block at @bytenr. If @gen is 0 then we do a
4964 * normal uptodate check of the eb, without checking the generation. If we have
4965 * to read the block we will not block on anything.
4966 */
4967void btrfs_readahead_tree_block(struct btrfs_fs_info *fs_info,
4968 u64 bytenr, u64 owner_root, u64 gen, int level)
4969{
4970 struct btrfs_tree_parent_check check = {
4971 .has_first_key = 0,
4972 .level = level,
4973 .transid = gen
4974 };
4975 struct extent_buffer *eb;
4976 int ret;
4977
4978 eb = btrfs_find_create_tree_block(fs_info, bytenr, owner_root, level);
4979 if (IS_ERR(eb))
4980 return;
4981
4982 if (btrfs_buffer_uptodate(eb, gen, 1)) {
4983 free_extent_buffer(eb);
4984 return;
4985 }
4986
4987 ret = read_extent_buffer_pages(eb, WAIT_NONE, 0, &check);
4988 if (ret < 0)
4989 free_extent_buffer_stale(eb);
4990 else
4991 free_extent_buffer(eb);
4992}
4993
4994/*
4995 * Readahead a node's child block.
4996 *
4997 * @node: parent node we're reading from
4998 * @slot: slot in the parent node for the child we want to read
4999 *
5000 * A helper for btrfs_readahead_tree_block, we simply read the bytenr pointed at
5001 * the slot in the node provided.
5002 */
5003void btrfs_readahead_node_child(struct extent_buffer *node, int slot)
5004{
5005 btrfs_readahead_tree_block(node->fs_info,
5006 btrfs_node_blockptr(node, slot),
5007 btrfs_header_owner(node),
5008 btrfs_node_ptr_generation(node, slot),
5009 btrfs_header_level(node) - 1);
5010}
1// SPDX-License-Identifier: GPL-2.0
2
3#include <linux/bitops.h>
4#include <linux/slab.h>
5#include <linux/bio.h>
6#include <linux/mm.h>
7#include <linux/pagemap.h>
8#include <linux/page-flags.h>
9#include <linux/sched/mm.h>
10#include <linux/spinlock.h>
11#include <linux/blkdev.h>
12#include <linux/swap.h>
13#include <linux/writeback.h>
14#include <linux/pagevec.h>
15#include <linux/prefetch.h>
16#include <linux/fsverity.h>
17#include "extent_io.h"
18#include "extent-io-tree.h"
19#include "extent_map.h"
20#include "ctree.h"
21#include "btrfs_inode.h"
22#include "bio.h"
23#include "locking.h"
24#include "backref.h"
25#include "disk-io.h"
26#include "subpage.h"
27#include "zoned.h"
28#include "block-group.h"
29#include "compression.h"
30#include "fs.h"
31#include "accessors.h"
32#include "file-item.h"
33#include "file.h"
34#include "dev-replace.h"
35#include "super.h"
36#include "transaction.h"
37
38static struct kmem_cache *extent_buffer_cache;
39
40#ifdef CONFIG_BTRFS_DEBUG
41static inline void btrfs_leak_debug_add_eb(struct extent_buffer *eb)
42{
43 struct btrfs_fs_info *fs_info = eb->fs_info;
44 unsigned long flags;
45
46 spin_lock_irqsave(&fs_info->eb_leak_lock, flags);
47 list_add(&eb->leak_list, &fs_info->allocated_ebs);
48 spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
49}
50
51static inline void btrfs_leak_debug_del_eb(struct extent_buffer *eb)
52{
53 struct btrfs_fs_info *fs_info = eb->fs_info;
54 unsigned long flags;
55
56 spin_lock_irqsave(&fs_info->eb_leak_lock, flags);
57 list_del(&eb->leak_list);
58 spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
59}
60
61void btrfs_extent_buffer_leak_debug_check(struct btrfs_fs_info *fs_info)
62{
63 struct extent_buffer *eb;
64 unsigned long flags;
65
66 /*
67 * If we didn't get into open_ctree our allocated_ebs will not be
68 * initialized, so just skip this.
69 */
70 if (!fs_info->allocated_ebs.next)
71 return;
72
73 WARN_ON(!list_empty(&fs_info->allocated_ebs));
74 spin_lock_irqsave(&fs_info->eb_leak_lock, flags);
75 while (!list_empty(&fs_info->allocated_ebs)) {
76 eb = list_first_entry(&fs_info->allocated_ebs,
77 struct extent_buffer, leak_list);
78 pr_err(
79 "BTRFS: buffer leak start %llu len %u refs %d bflags %lu owner %llu\n",
80 eb->start, eb->len, atomic_read(&eb->refs), eb->bflags,
81 btrfs_header_owner(eb));
82 list_del(&eb->leak_list);
83 WARN_ON_ONCE(1);
84 kmem_cache_free(extent_buffer_cache, eb);
85 }
86 spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
87}
88#else
89#define btrfs_leak_debug_add_eb(eb) do {} while (0)
90#define btrfs_leak_debug_del_eb(eb) do {} while (0)
91#endif
92
93/*
94 * Structure to record info about the bio being assembled, and other info like
95 * how many bytes are there before stripe/ordered extent boundary.
96 */
97struct btrfs_bio_ctrl {
98 struct btrfs_bio *bbio;
99 enum btrfs_compression_type compress_type;
100 u32 len_to_oe_boundary;
101 blk_opf_t opf;
102 btrfs_bio_end_io_t end_io_func;
103 struct writeback_control *wbc;
104
105 /*
106 * The sectors of the page which are going to be submitted by
107 * extent_writepage_io().
108 * This is to avoid touching ranges covered by compression/inline.
109 */
110 unsigned long submit_bitmap;
111};
112
113static void submit_one_bio(struct btrfs_bio_ctrl *bio_ctrl)
114{
115 struct btrfs_bio *bbio = bio_ctrl->bbio;
116
117 if (!bbio)
118 return;
119
120 /* Caller should ensure the bio has at least some range added */
121 ASSERT(bbio->bio.bi_iter.bi_size);
122
123 if (btrfs_op(&bbio->bio) == BTRFS_MAP_READ &&
124 bio_ctrl->compress_type != BTRFS_COMPRESS_NONE)
125 btrfs_submit_compressed_read(bbio);
126 else
127 btrfs_submit_bbio(bbio, 0);
128
129 /* The bbio is owned by the end_io handler now */
130 bio_ctrl->bbio = NULL;
131}
132
133/*
134 * Submit or fail the current bio in the bio_ctrl structure.
135 */
136static void submit_write_bio(struct btrfs_bio_ctrl *bio_ctrl, int ret)
137{
138 struct btrfs_bio *bbio = bio_ctrl->bbio;
139
140 if (!bbio)
141 return;
142
143 if (ret) {
144 ASSERT(ret < 0);
145 btrfs_bio_end_io(bbio, errno_to_blk_status(ret));
146 /* The bio is owned by the end_io handler now */
147 bio_ctrl->bbio = NULL;
148 } else {
149 submit_one_bio(bio_ctrl);
150 }
151}
152
153int __init extent_buffer_init_cachep(void)
154{
155 extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
156 sizeof(struct extent_buffer), 0, 0,
157 NULL);
158 if (!extent_buffer_cache)
159 return -ENOMEM;
160
161 return 0;
162}
163
164void __cold extent_buffer_free_cachep(void)
165{
166 /*
167 * Make sure all delayed rcu free are flushed before we
168 * destroy caches.
169 */
170 rcu_barrier();
171 kmem_cache_destroy(extent_buffer_cache);
172}
173
174static void process_one_folio(struct btrfs_fs_info *fs_info,
175 struct folio *folio, const struct folio *locked_folio,
176 unsigned long page_ops, u64 start, u64 end)
177{
178 u32 len;
179
180 ASSERT(end + 1 - start != 0 && end + 1 - start < U32_MAX);
181 len = end + 1 - start;
182
183 if (page_ops & PAGE_SET_ORDERED)
184 btrfs_folio_clamp_set_ordered(fs_info, folio, start, len);
185 if (page_ops & PAGE_START_WRITEBACK) {
186 btrfs_folio_clamp_clear_dirty(fs_info, folio, start, len);
187 btrfs_folio_clamp_set_writeback(fs_info, folio, start, len);
188 }
189 if (page_ops & PAGE_END_WRITEBACK)
190 btrfs_folio_clamp_clear_writeback(fs_info, folio, start, len);
191
192 if (folio != locked_folio && (page_ops & PAGE_UNLOCK))
193 btrfs_folio_end_lock(fs_info, folio, start, len);
194}
195
196static void __process_folios_contig(struct address_space *mapping,
197 const struct folio *locked_folio, u64 start,
198 u64 end, unsigned long page_ops)
199{
200 struct btrfs_fs_info *fs_info = inode_to_fs_info(mapping->host);
201 pgoff_t start_index = start >> PAGE_SHIFT;
202 pgoff_t end_index = end >> PAGE_SHIFT;
203 pgoff_t index = start_index;
204 struct folio_batch fbatch;
205 int i;
206
207 folio_batch_init(&fbatch);
208 while (index <= end_index) {
209 int found_folios;
210
211 found_folios = filemap_get_folios_contig(mapping, &index,
212 end_index, &fbatch);
213 for (i = 0; i < found_folios; i++) {
214 struct folio *folio = fbatch.folios[i];
215
216 process_one_folio(fs_info, folio, locked_folio,
217 page_ops, start, end);
218 }
219 folio_batch_release(&fbatch);
220 cond_resched();
221 }
222}
223
224static noinline void __unlock_for_delalloc(const struct inode *inode,
225 const struct folio *locked_folio,
226 u64 start, u64 end)
227{
228 unsigned long index = start >> PAGE_SHIFT;
229 unsigned long end_index = end >> PAGE_SHIFT;
230
231 ASSERT(locked_folio);
232 if (index == locked_folio->index && end_index == index)
233 return;
234
235 __process_folios_contig(inode->i_mapping, locked_folio, start, end,
236 PAGE_UNLOCK);
237}
238
239static noinline int lock_delalloc_folios(struct inode *inode,
240 const struct folio *locked_folio,
241 u64 start, u64 end)
242{
243 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
244 struct address_space *mapping = inode->i_mapping;
245 pgoff_t start_index = start >> PAGE_SHIFT;
246 pgoff_t end_index = end >> PAGE_SHIFT;
247 pgoff_t index = start_index;
248 u64 processed_end = start;
249 struct folio_batch fbatch;
250
251 if (index == locked_folio->index && index == end_index)
252 return 0;
253
254 folio_batch_init(&fbatch);
255 while (index <= end_index) {
256 unsigned int found_folios, i;
257
258 found_folios = filemap_get_folios_contig(mapping, &index,
259 end_index, &fbatch);
260 if (found_folios == 0)
261 goto out;
262
263 for (i = 0; i < found_folios; i++) {
264 struct folio *folio = fbatch.folios[i];
265 u64 range_start;
266 u32 range_len;
267
268 if (folio == locked_folio)
269 continue;
270
271 folio_lock(folio);
272 if (!folio_test_dirty(folio) || folio->mapping != mapping) {
273 folio_unlock(folio);
274 goto out;
275 }
276 range_start = max_t(u64, folio_pos(folio), start);
277 range_len = min_t(u64, folio_pos(folio) + folio_size(folio),
278 end + 1) - range_start;
279 btrfs_folio_set_lock(fs_info, folio, range_start, range_len);
280
281 processed_end = range_start + range_len - 1;
282 }
283 folio_batch_release(&fbatch);
284 cond_resched();
285 }
286
287 return 0;
288out:
289 folio_batch_release(&fbatch);
290 if (processed_end > start)
291 __unlock_for_delalloc(inode, locked_folio, start,
292 processed_end);
293 return -EAGAIN;
294}
295
296/*
297 * Find and lock a contiguous range of bytes in the file marked as delalloc, no
298 * more than @max_bytes.
299 *
300 * @start: The original start bytenr to search.
301 * Will store the extent range start bytenr.
302 * @end: The original end bytenr of the search range
303 * Will store the extent range end bytenr.
304 *
305 * Return true if we find a delalloc range which starts inside the original
306 * range, and @start/@end will store the delalloc range start/end.
307 *
308 * Return false if we can't find any delalloc range which starts inside the
309 * original range, and @start/@end will be the non-delalloc range start/end.
310 */
311EXPORT_FOR_TESTS
312noinline_for_stack bool find_lock_delalloc_range(struct inode *inode,
313 struct folio *locked_folio,
314 u64 *start, u64 *end)
315{
316 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
317 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
318 const u64 orig_start = *start;
319 const u64 orig_end = *end;
320 /* The sanity tests may not set a valid fs_info. */
321 u64 max_bytes = fs_info ? fs_info->max_extent_size : BTRFS_MAX_EXTENT_SIZE;
322 u64 delalloc_start;
323 u64 delalloc_end;
324 bool found;
325 struct extent_state *cached_state = NULL;
326 int ret;
327 int loops = 0;
328
329 /* Caller should pass a valid @end to indicate the search range end */
330 ASSERT(orig_end > orig_start);
331
332 /* The range should at least cover part of the folio */
333 ASSERT(!(orig_start >= folio_pos(locked_folio) + folio_size(locked_folio) ||
334 orig_end <= folio_pos(locked_folio)));
335again:
336 /* step one, find a bunch of delalloc bytes starting at start */
337 delalloc_start = *start;
338 delalloc_end = 0;
339 found = btrfs_find_delalloc_range(tree, &delalloc_start, &delalloc_end,
340 max_bytes, &cached_state);
341 if (!found || delalloc_end <= *start || delalloc_start > orig_end) {
342 *start = delalloc_start;
343
344 /* @delalloc_end can be -1, never go beyond @orig_end */
345 *end = min(delalloc_end, orig_end);
346 free_extent_state(cached_state);
347 return false;
348 }
349
350 /*
351 * start comes from the offset of locked_folio. We have to lock
352 * folios in order, so we can't process delalloc bytes before
353 * locked_folio
354 */
355 if (delalloc_start < *start)
356 delalloc_start = *start;
357
358 /*
359 * make sure to limit the number of folios we try to lock down
360 */
361 if (delalloc_end + 1 - delalloc_start > max_bytes)
362 delalloc_end = delalloc_start + max_bytes - 1;
363
364 /* step two, lock all the folioss after the folios that has start */
365 ret = lock_delalloc_folios(inode, locked_folio, delalloc_start,
366 delalloc_end);
367 ASSERT(!ret || ret == -EAGAIN);
368 if (ret == -EAGAIN) {
369 /* some of the folios are gone, lets avoid looping by
370 * shortening the size of the delalloc range we're searching
371 */
372 free_extent_state(cached_state);
373 cached_state = NULL;
374 if (!loops) {
375 max_bytes = PAGE_SIZE;
376 loops = 1;
377 goto again;
378 } else {
379 found = false;
380 goto out_failed;
381 }
382 }
383
384 /* step three, lock the state bits for the whole range */
385 lock_extent(tree, delalloc_start, delalloc_end, &cached_state);
386
387 /* then test to make sure it is all still delalloc */
388 ret = test_range_bit(tree, delalloc_start, delalloc_end,
389 EXTENT_DELALLOC, cached_state);
390
391 unlock_extent(tree, delalloc_start, delalloc_end, &cached_state);
392 if (!ret) {
393 __unlock_for_delalloc(inode, locked_folio, delalloc_start,
394 delalloc_end);
395 cond_resched();
396 goto again;
397 }
398 *start = delalloc_start;
399 *end = delalloc_end;
400out_failed:
401 return found;
402}
403
404void extent_clear_unlock_delalloc(struct btrfs_inode *inode, u64 start, u64 end,
405 const struct folio *locked_folio,
406 struct extent_state **cached,
407 u32 clear_bits, unsigned long page_ops)
408{
409 clear_extent_bit(&inode->io_tree, start, end, clear_bits, cached);
410
411 __process_folios_contig(inode->vfs_inode.i_mapping, locked_folio, start,
412 end, page_ops);
413}
414
415static bool btrfs_verify_folio(struct folio *folio, u64 start, u32 len)
416{
417 struct btrfs_fs_info *fs_info = folio_to_fs_info(folio);
418
419 if (!fsverity_active(folio->mapping->host) ||
420 btrfs_folio_test_uptodate(fs_info, folio, start, len) ||
421 start >= i_size_read(folio->mapping->host))
422 return true;
423 return fsverity_verify_folio(folio);
424}
425
426static void end_folio_read(struct folio *folio, bool uptodate, u64 start, u32 len)
427{
428 struct btrfs_fs_info *fs_info = folio_to_fs_info(folio);
429
430 ASSERT(folio_pos(folio) <= start &&
431 start + len <= folio_pos(folio) + PAGE_SIZE);
432
433 if (uptodate && btrfs_verify_folio(folio, start, len))
434 btrfs_folio_set_uptodate(fs_info, folio, start, len);
435 else
436 btrfs_folio_clear_uptodate(fs_info, folio, start, len);
437
438 if (!btrfs_is_subpage(fs_info, folio->mapping))
439 folio_unlock(folio);
440 else
441 btrfs_folio_end_lock(fs_info, folio, start, len);
442}
443
444/*
445 * After a write IO is done, we need to:
446 *
447 * - clear the uptodate bits on error
448 * - clear the writeback bits in the extent tree for the range
449 * - filio_end_writeback() if there is no more pending io for the folio
450 *
451 * Scheduling is not allowed, so the extent state tree is expected
452 * to have one and only one object corresponding to this IO.
453 */
454static void end_bbio_data_write(struct btrfs_bio *bbio)
455{
456 struct btrfs_fs_info *fs_info = bbio->fs_info;
457 struct bio *bio = &bbio->bio;
458 int error = blk_status_to_errno(bio->bi_status);
459 struct folio_iter fi;
460 const u32 sectorsize = fs_info->sectorsize;
461
462 ASSERT(!bio_flagged(bio, BIO_CLONED));
463 bio_for_each_folio_all(fi, bio) {
464 struct folio *folio = fi.folio;
465 u64 start = folio_pos(folio) + fi.offset;
466 u32 len = fi.length;
467
468 /* Only order 0 (single page) folios are allowed for data. */
469 ASSERT(folio_order(folio) == 0);
470
471 /* Our read/write should always be sector aligned. */
472 if (!IS_ALIGNED(fi.offset, sectorsize))
473 btrfs_err(fs_info,
474 "partial page write in btrfs with offset %zu and length %zu",
475 fi.offset, fi.length);
476 else if (!IS_ALIGNED(fi.length, sectorsize))
477 btrfs_info(fs_info,
478 "incomplete page write with offset %zu and length %zu",
479 fi.offset, fi.length);
480
481 btrfs_finish_ordered_extent(bbio->ordered, folio, start, len,
482 !error);
483 if (error)
484 mapping_set_error(folio->mapping, error);
485 btrfs_folio_clear_writeback(fs_info, folio, start, len);
486 }
487
488 bio_put(bio);
489}
490
491static void begin_folio_read(struct btrfs_fs_info *fs_info, struct folio *folio)
492{
493 ASSERT(folio_test_locked(folio));
494 if (!btrfs_is_subpage(fs_info, folio->mapping))
495 return;
496
497 ASSERT(folio_test_private(folio));
498 btrfs_folio_set_lock(fs_info, folio, folio_pos(folio), PAGE_SIZE);
499}
500
501/*
502 * After a data read IO is done, we need to:
503 *
504 * - clear the uptodate bits on error
505 * - set the uptodate bits if things worked
506 * - set the folio up to date if all extents in the tree are uptodate
507 * - clear the lock bit in the extent tree
508 * - unlock the folio if there are no other extents locked for it
509 *
510 * Scheduling is not allowed, so the extent state tree is expected
511 * to have one and only one object corresponding to this IO.
512 */
513static void end_bbio_data_read(struct btrfs_bio *bbio)
514{
515 struct btrfs_fs_info *fs_info = bbio->fs_info;
516 struct bio *bio = &bbio->bio;
517 struct folio_iter fi;
518 const u32 sectorsize = fs_info->sectorsize;
519
520 ASSERT(!bio_flagged(bio, BIO_CLONED));
521 bio_for_each_folio_all(fi, &bbio->bio) {
522 bool uptodate = !bio->bi_status;
523 struct folio *folio = fi.folio;
524 struct inode *inode = folio->mapping->host;
525 u64 start;
526 u64 end;
527 u32 len;
528
529 /* For now only order 0 folios are supported for data. */
530 ASSERT(folio_order(folio) == 0);
531 btrfs_debug(fs_info,
532 "%s: bi_sector=%llu, err=%d, mirror=%u",
533 __func__, bio->bi_iter.bi_sector, bio->bi_status,
534 bbio->mirror_num);
535
536 /*
537 * We always issue full-sector reads, but if some block in a
538 * folio fails to read, blk_update_request() will advance
539 * bv_offset and adjust bv_len to compensate. Print a warning
540 * for unaligned offsets, and an error if they don't add up to
541 * a full sector.
542 */
543 if (!IS_ALIGNED(fi.offset, sectorsize))
544 btrfs_err(fs_info,
545 "partial page read in btrfs with offset %zu and length %zu",
546 fi.offset, fi.length);
547 else if (!IS_ALIGNED(fi.offset + fi.length, sectorsize))
548 btrfs_info(fs_info,
549 "incomplete page read with offset %zu and length %zu",
550 fi.offset, fi.length);
551
552 start = folio_pos(folio) + fi.offset;
553 end = start + fi.length - 1;
554 len = fi.length;
555
556 if (likely(uptodate)) {
557 loff_t i_size = i_size_read(inode);
558 pgoff_t end_index = i_size >> folio_shift(folio);
559
560 /*
561 * Zero out the remaining part if this range straddles
562 * i_size.
563 *
564 * Here we should only zero the range inside the folio,
565 * not touch anything else.
566 *
567 * NOTE: i_size is exclusive while end is inclusive.
568 */
569 if (folio_index(folio) == end_index && i_size <= end) {
570 u32 zero_start = max(offset_in_folio(folio, i_size),
571 offset_in_folio(folio, start));
572 u32 zero_len = offset_in_folio(folio, end) + 1 -
573 zero_start;
574
575 folio_zero_range(folio, zero_start, zero_len);
576 }
577 }
578
579 /* Update page status and unlock. */
580 end_folio_read(folio, uptodate, start, len);
581 }
582 bio_put(bio);
583}
584
585/*
586 * Populate every free slot in a provided array with folios using GFP_NOFS.
587 *
588 * @nr_folios: number of folios to allocate
589 * @folio_array: the array to fill with folios; any existing non-NULL entries in
590 * the array will be skipped
591 *
592 * Return: 0 if all folios were able to be allocated;
593 * -ENOMEM otherwise, the partially allocated folios would be freed and
594 * the array slots zeroed
595 */
596int btrfs_alloc_folio_array(unsigned int nr_folios, struct folio **folio_array)
597{
598 for (int i = 0; i < nr_folios; i++) {
599 if (folio_array[i])
600 continue;
601 folio_array[i] = folio_alloc(GFP_NOFS, 0);
602 if (!folio_array[i])
603 goto error;
604 }
605 return 0;
606error:
607 for (int i = 0; i < nr_folios; i++) {
608 if (folio_array[i])
609 folio_put(folio_array[i]);
610 }
611 return -ENOMEM;
612}
613
614/*
615 * Populate every free slot in a provided array with pages, using GFP_NOFS.
616 *
617 * @nr_pages: number of pages to allocate
618 * @page_array: the array to fill with pages; any existing non-null entries in
619 * the array will be skipped
620 * @nofail: whether using __GFP_NOFAIL flag
621 *
622 * Return: 0 if all pages were able to be allocated;
623 * -ENOMEM otherwise, the partially allocated pages would be freed and
624 * the array slots zeroed
625 */
626int btrfs_alloc_page_array(unsigned int nr_pages, struct page **page_array,
627 bool nofail)
628{
629 const gfp_t gfp = nofail ? (GFP_NOFS | __GFP_NOFAIL) : GFP_NOFS;
630 unsigned int allocated;
631
632 for (allocated = 0; allocated < nr_pages;) {
633 unsigned int last = allocated;
634
635 allocated = alloc_pages_bulk_array(gfp, nr_pages, page_array);
636 if (unlikely(allocated == last)) {
637 /* No progress, fail and do cleanup. */
638 for (int i = 0; i < allocated; i++) {
639 __free_page(page_array[i]);
640 page_array[i] = NULL;
641 }
642 return -ENOMEM;
643 }
644 }
645 return 0;
646}
647
648/*
649 * Populate needed folios for the extent buffer.
650 *
651 * For now, the folios populated are always in order 0 (aka, single page).
652 */
653static int alloc_eb_folio_array(struct extent_buffer *eb, bool nofail)
654{
655 struct page *page_array[INLINE_EXTENT_BUFFER_PAGES] = { 0 };
656 int num_pages = num_extent_pages(eb);
657 int ret;
658
659 ret = btrfs_alloc_page_array(num_pages, page_array, nofail);
660 if (ret < 0)
661 return ret;
662
663 for (int i = 0; i < num_pages; i++)
664 eb->folios[i] = page_folio(page_array[i]);
665 eb->folio_size = PAGE_SIZE;
666 eb->folio_shift = PAGE_SHIFT;
667 return 0;
668}
669
670static bool btrfs_bio_is_contig(struct btrfs_bio_ctrl *bio_ctrl,
671 struct folio *folio, u64 disk_bytenr,
672 unsigned int pg_offset)
673{
674 struct bio *bio = &bio_ctrl->bbio->bio;
675 struct bio_vec *bvec = bio_last_bvec_all(bio);
676 const sector_t sector = disk_bytenr >> SECTOR_SHIFT;
677 struct folio *bv_folio = page_folio(bvec->bv_page);
678
679 if (bio_ctrl->compress_type != BTRFS_COMPRESS_NONE) {
680 /*
681 * For compression, all IO should have its logical bytenr set
682 * to the starting bytenr of the compressed extent.
683 */
684 return bio->bi_iter.bi_sector == sector;
685 }
686
687 /*
688 * The contig check requires the following conditions to be met:
689 *
690 * 1) The folios are belonging to the same inode
691 * This is implied by the call chain.
692 *
693 * 2) The range has adjacent logical bytenr
694 *
695 * 3) The range has adjacent file offset
696 * This is required for the usage of btrfs_bio->file_offset.
697 */
698 return bio_end_sector(bio) == sector &&
699 folio_pos(bv_folio) + bvec->bv_offset + bvec->bv_len ==
700 folio_pos(folio) + pg_offset;
701}
702
703static void alloc_new_bio(struct btrfs_inode *inode,
704 struct btrfs_bio_ctrl *bio_ctrl,
705 u64 disk_bytenr, u64 file_offset)
706{
707 struct btrfs_fs_info *fs_info = inode->root->fs_info;
708 struct btrfs_bio *bbio;
709
710 bbio = btrfs_bio_alloc(BIO_MAX_VECS, bio_ctrl->opf, fs_info,
711 bio_ctrl->end_io_func, NULL);
712 bbio->bio.bi_iter.bi_sector = disk_bytenr >> SECTOR_SHIFT;
713 bbio->inode = inode;
714 bbio->file_offset = file_offset;
715 bio_ctrl->bbio = bbio;
716 bio_ctrl->len_to_oe_boundary = U32_MAX;
717
718 /* Limit data write bios to the ordered boundary. */
719 if (bio_ctrl->wbc) {
720 struct btrfs_ordered_extent *ordered;
721
722 ordered = btrfs_lookup_ordered_extent(inode, file_offset);
723 if (ordered) {
724 bio_ctrl->len_to_oe_boundary = min_t(u32, U32_MAX,
725 ordered->file_offset +
726 ordered->disk_num_bytes - file_offset);
727 bbio->ordered = ordered;
728 }
729
730 /*
731 * Pick the last added device to support cgroup writeback. For
732 * multi-device file systems this means blk-cgroup policies have
733 * to always be set on the last added/replaced device.
734 * This is a bit odd but has been like that for a long time.
735 */
736 bio_set_dev(&bbio->bio, fs_info->fs_devices->latest_dev->bdev);
737 wbc_init_bio(bio_ctrl->wbc, &bbio->bio);
738 }
739}
740
741/*
742 * @disk_bytenr: logical bytenr where the write will be
743 * @page: page to add to the bio
744 * @size: portion of page that we want to write to
745 * @pg_offset: offset of the new bio or to check whether we are adding
746 * a contiguous page to the previous one
747 *
748 * The will either add the page into the existing @bio_ctrl->bbio, or allocate a
749 * new one in @bio_ctrl->bbio.
750 * The mirror number for this IO should already be initizlied in
751 * @bio_ctrl->mirror_num.
752 */
753static void submit_extent_folio(struct btrfs_bio_ctrl *bio_ctrl,
754 u64 disk_bytenr, struct folio *folio,
755 size_t size, unsigned long pg_offset)
756{
757 struct btrfs_inode *inode = folio_to_inode(folio);
758
759 ASSERT(pg_offset + size <= PAGE_SIZE);
760 ASSERT(bio_ctrl->end_io_func);
761
762 if (bio_ctrl->bbio &&
763 !btrfs_bio_is_contig(bio_ctrl, folio, disk_bytenr, pg_offset))
764 submit_one_bio(bio_ctrl);
765
766 do {
767 u32 len = size;
768
769 /* Allocate new bio if needed */
770 if (!bio_ctrl->bbio) {
771 alloc_new_bio(inode, bio_ctrl, disk_bytenr,
772 folio_pos(folio) + pg_offset);
773 }
774
775 /* Cap to the current ordered extent boundary if there is one. */
776 if (len > bio_ctrl->len_to_oe_boundary) {
777 ASSERT(bio_ctrl->compress_type == BTRFS_COMPRESS_NONE);
778 ASSERT(is_data_inode(inode));
779 len = bio_ctrl->len_to_oe_boundary;
780 }
781
782 if (!bio_add_folio(&bio_ctrl->bbio->bio, folio, len, pg_offset)) {
783 /* bio full: move on to a new one */
784 submit_one_bio(bio_ctrl);
785 continue;
786 }
787
788 if (bio_ctrl->wbc)
789 wbc_account_cgroup_owner(bio_ctrl->wbc, folio,
790 len);
791
792 size -= len;
793 pg_offset += len;
794 disk_bytenr += len;
795
796 /*
797 * len_to_oe_boundary defaults to U32_MAX, which isn't folio or
798 * sector aligned. alloc_new_bio() then sets it to the end of
799 * our ordered extent for writes into zoned devices.
800 *
801 * When len_to_oe_boundary is tracking an ordered extent, we
802 * trust the ordered extent code to align things properly, and
803 * the check above to cap our write to the ordered extent
804 * boundary is correct.
805 *
806 * When len_to_oe_boundary is U32_MAX, the cap above would
807 * result in a 4095 byte IO for the last folio right before
808 * we hit the bio limit of UINT_MAX. bio_add_folio() has all
809 * the checks required to make sure we don't overflow the bio,
810 * and we should just ignore len_to_oe_boundary completely
811 * unless we're using it to track an ordered extent.
812 *
813 * It's pretty hard to make a bio sized U32_MAX, but it can
814 * happen when the page cache is able to feed us contiguous
815 * folios for large extents.
816 */
817 if (bio_ctrl->len_to_oe_boundary != U32_MAX)
818 bio_ctrl->len_to_oe_boundary -= len;
819
820 /* Ordered extent boundary: move on to a new bio. */
821 if (bio_ctrl->len_to_oe_boundary == 0)
822 submit_one_bio(bio_ctrl);
823 } while (size);
824}
825
826static int attach_extent_buffer_folio(struct extent_buffer *eb,
827 struct folio *folio,
828 struct btrfs_subpage *prealloc)
829{
830 struct btrfs_fs_info *fs_info = eb->fs_info;
831 int ret = 0;
832
833 /*
834 * If the page is mapped to btree inode, we should hold the private
835 * lock to prevent race.
836 * For cloned or dummy extent buffers, their pages are not mapped and
837 * will not race with any other ebs.
838 */
839 if (folio->mapping)
840 lockdep_assert_held(&folio->mapping->i_private_lock);
841
842 if (fs_info->nodesize >= PAGE_SIZE) {
843 if (!folio_test_private(folio))
844 folio_attach_private(folio, eb);
845 else
846 WARN_ON(folio_get_private(folio) != eb);
847 return 0;
848 }
849
850 /* Already mapped, just free prealloc */
851 if (folio_test_private(folio)) {
852 btrfs_free_subpage(prealloc);
853 return 0;
854 }
855
856 if (prealloc)
857 /* Has preallocated memory for subpage */
858 folio_attach_private(folio, prealloc);
859 else
860 /* Do new allocation to attach subpage */
861 ret = btrfs_attach_subpage(fs_info, folio, BTRFS_SUBPAGE_METADATA);
862 return ret;
863}
864
865int set_page_extent_mapped(struct page *page)
866{
867 return set_folio_extent_mapped(page_folio(page));
868}
869
870int set_folio_extent_mapped(struct folio *folio)
871{
872 struct btrfs_fs_info *fs_info;
873
874 ASSERT(folio->mapping);
875
876 if (folio_test_private(folio))
877 return 0;
878
879 fs_info = folio_to_fs_info(folio);
880
881 if (btrfs_is_subpage(fs_info, folio->mapping))
882 return btrfs_attach_subpage(fs_info, folio, BTRFS_SUBPAGE_DATA);
883
884 folio_attach_private(folio, (void *)EXTENT_FOLIO_PRIVATE);
885 return 0;
886}
887
888void clear_folio_extent_mapped(struct folio *folio)
889{
890 struct btrfs_fs_info *fs_info;
891
892 ASSERT(folio->mapping);
893
894 if (!folio_test_private(folio))
895 return;
896
897 fs_info = folio_to_fs_info(folio);
898 if (btrfs_is_subpage(fs_info, folio->mapping))
899 return btrfs_detach_subpage(fs_info, folio);
900
901 folio_detach_private(folio);
902}
903
904static struct extent_map *get_extent_map(struct btrfs_inode *inode,
905 struct folio *folio, u64 start,
906 u64 len, struct extent_map **em_cached)
907{
908 struct extent_map *em;
909
910 ASSERT(em_cached);
911
912 if (*em_cached) {
913 em = *em_cached;
914 if (extent_map_in_tree(em) && start >= em->start &&
915 start < extent_map_end(em)) {
916 refcount_inc(&em->refs);
917 return em;
918 }
919
920 free_extent_map(em);
921 *em_cached = NULL;
922 }
923
924 em = btrfs_get_extent(inode, folio, start, len);
925 if (!IS_ERR(em)) {
926 BUG_ON(*em_cached);
927 refcount_inc(&em->refs);
928 *em_cached = em;
929 }
930
931 return em;
932}
933/*
934 * basic readpage implementation. Locked extent state structs are inserted
935 * into the tree that are removed when the IO is done (by the end_io
936 * handlers)
937 * XXX JDM: This needs looking at to ensure proper page locking
938 * return 0 on success, otherwise return error
939 */
940static int btrfs_do_readpage(struct folio *folio, struct extent_map **em_cached,
941 struct btrfs_bio_ctrl *bio_ctrl, u64 *prev_em_start)
942{
943 struct inode *inode = folio->mapping->host;
944 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
945 u64 start = folio_pos(folio);
946 const u64 end = start + PAGE_SIZE - 1;
947 u64 cur = start;
948 u64 extent_offset;
949 u64 last_byte = i_size_read(inode);
950 u64 block_start;
951 struct extent_map *em;
952 int ret = 0;
953 size_t pg_offset = 0;
954 size_t iosize;
955 size_t blocksize = fs_info->sectorsize;
956
957 ret = set_folio_extent_mapped(folio);
958 if (ret < 0) {
959 folio_unlock(folio);
960 return ret;
961 }
962
963 if (folio->index == last_byte >> folio_shift(folio)) {
964 size_t zero_offset = offset_in_folio(folio, last_byte);
965
966 if (zero_offset) {
967 iosize = folio_size(folio) - zero_offset;
968 folio_zero_range(folio, zero_offset, iosize);
969 }
970 }
971 bio_ctrl->end_io_func = end_bbio_data_read;
972 begin_folio_read(fs_info, folio);
973 while (cur <= end) {
974 enum btrfs_compression_type compress_type = BTRFS_COMPRESS_NONE;
975 bool force_bio_submit = false;
976 u64 disk_bytenr;
977
978 ASSERT(IS_ALIGNED(cur, fs_info->sectorsize));
979 if (cur >= last_byte) {
980 iosize = folio_size(folio) - pg_offset;
981 folio_zero_range(folio, pg_offset, iosize);
982 end_folio_read(folio, true, cur, iosize);
983 break;
984 }
985 em = get_extent_map(BTRFS_I(inode), folio, cur, end - cur + 1, em_cached);
986 if (IS_ERR(em)) {
987 end_folio_read(folio, false, cur, end + 1 - cur);
988 return PTR_ERR(em);
989 }
990 extent_offset = cur - em->start;
991 BUG_ON(extent_map_end(em) <= cur);
992 BUG_ON(end < cur);
993
994 compress_type = extent_map_compression(em);
995
996 iosize = min(extent_map_end(em) - cur, end - cur + 1);
997 iosize = ALIGN(iosize, blocksize);
998 if (compress_type != BTRFS_COMPRESS_NONE)
999 disk_bytenr = em->disk_bytenr;
1000 else
1001 disk_bytenr = extent_map_block_start(em) + extent_offset;
1002 block_start = extent_map_block_start(em);
1003 if (em->flags & EXTENT_FLAG_PREALLOC)
1004 block_start = EXTENT_MAP_HOLE;
1005
1006 /*
1007 * If we have a file range that points to a compressed extent
1008 * and it's followed by a consecutive file range that points
1009 * to the same compressed extent (possibly with a different
1010 * offset and/or length, so it either points to the whole extent
1011 * or only part of it), we must make sure we do not submit a
1012 * single bio to populate the folios for the 2 ranges because
1013 * this makes the compressed extent read zero out the folios
1014 * belonging to the 2nd range. Imagine the following scenario:
1015 *
1016 * File layout
1017 * [0 - 8K] [8K - 24K]
1018 * | |
1019 * | |
1020 * points to extent X, points to extent X,
1021 * offset 4K, length of 8K offset 0, length 16K
1022 *
1023 * [extent X, compressed length = 4K uncompressed length = 16K]
1024 *
1025 * If the bio to read the compressed extent covers both ranges,
1026 * it will decompress extent X into the folios belonging to the
1027 * first range and then it will stop, zeroing out the remaining
1028 * folios that belong to the other range that points to extent X.
1029 * So here we make sure we submit 2 bios, one for the first
1030 * range and another one for the third range. Both will target
1031 * the same physical extent from disk, but we can't currently
1032 * make the compressed bio endio callback populate the folios
1033 * for both ranges because each compressed bio is tightly
1034 * coupled with a single extent map, and each range can have
1035 * an extent map with a different offset value relative to the
1036 * uncompressed data of our extent and different lengths. This
1037 * is a corner case so we prioritize correctness over
1038 * non-optimal behavior (submitting 2 bios for the same extent).
1039 */
1040 if (compress_type != BTRFS_COMPRESS_NONE &&
1041 prev_em_start && *prev_em_start != (u64)-1 &&
1042 *prev_em_start != em->start)
1043 force_bio_submit = true;
1044
1045 if (prev_em_start)
1046 *prev_em_start = em->start;
1047
1048 free_extent_map(em);
1049 em = NULL;
1050
1051 /* we've found a hole, just zero and go on */
1052 if (block_start == EXTENT_MAP_HOLE) {
1053 folio_zero_range(folio, pg_offset, iosize);
1054
1055 end_folio_read(folio, true, cur, iosize);
1056 cur = cur + iosize;
1057 pg_offset += iosize;
1058 continue;
1059 }
1060 /* the get_extent function already copied into the folio */
1061 if (block_start == EXTENT_MAP_INLINE) {
1062 end_folio_read(folio, true, cur, iosize);
1063 cur = cur + iosize;
1064 pg_offset += iosize;
1065 continue;
1066 }
1067
1068 if (bio_ctrl->compress_type != compress_type) {
1069 submit_one_bio(bio_ctrl);
1070 bio_ctrl->compress_type = compress_type;
1071 }
1072
1073 if (force_bio_submit)
1074 submit_one_bio(bio_ctrl);
1075 submit_extent_folio(bio_ctrl, disk_bytenr, folio, iosize,
1076 pg_offset);
1077 cur = cur + iosize;
1078 pg_offset += iosize;
1079 }
1080
1081 return 0;
1082}
1083
1084int btrfs_read_folio(struct file *file, struct folio *folio)
1085{
1086 struct btrfs_inode *inode = folio_to_inode(folio);
1087 const u64 start = folio_pos(folio);
1088 const u64 end = start + folio_size(folio) - 1;
1089 struct extent_state *cached_state = NULL;
1090 struct btrfs_bio_ctrl bio_ctrl = { .opf = REQ_OP_READ };
1091 struct extent_map *em_cached = NULL;
1092 int ret;
1093
1094 btrfs_lock_and_flush_ordered_range(inode, start, end, &cached_state);
1095 ret = btrfs_do_readpage(folio, &em_cached, &bio_ctrl, NULL);
1096 unlock_extent(&inode->io_tree, start, end, &cached_state);
1097
1098 free_extent_map(em_cached);
1099
1100 /*
1101 * If btrfs_do_readpage() failed we will want to submit the assembled
1102 * bio to do the cleanup.
1103 */
1104 submit_one_bio(&bio_ctrl);
1105 return ret;
1106}
1107
1108static void set_delalloc_bitmap(struct folio *folio, unsigned long *delalloc_bitmap,
1109 u64 start, u32 len)
1110{
1111 struct btrfs_fs_info *fs_info = folio_to_fs_info(folio);
1112 const u64 folio_start = folio_pos(folio);
1113 unsigned int start_bit;
1114 unsigned int nbits;
1115
1116 ASSERT(start >= folio_start && start + len <= folio_start + PAGE_SIZE);
1117 start_bit = (start - folio_start) >> fs_info->sectorsize_bits;
1118 nbits = len >> fs_info->sectorsize_bits;
1119 ASSERT(bitmap_test_range_all_zero(delalloc_bitmap, start_bit, nbits));
1120 bitmap_set(delalloc_bitmap, start_bit, nbits);
1121}
1122
1123static bool find_next_delalloc_bitmap(struct folio *folio,
1124 unsigned long *delalloc_bitmap, u64 start,
1125 u64 *found_start, u32 *found_len)
1126{
1127 struct btrfs_fs_info *fs_info = folio_to_fs_info(folio);
1128 const u64 folio_start = folio_pos(folio);
1129 const unsigned int bitmap_size = fs_info->sectors_per_page;
1130 unsigned int start_bit;
1131 unsigned int first_zero;
1132 unsigned int first_set;
1133
1134 ASSERT(start >= folio_start && start < folio_start + PAGE_SIZE);
1135
1136 start_bit = (start - folio_start) >> fs_info->sectorsize_bits;
1137 first_set = find_next_bit(delalloc_bitmap, bitmap_size, start_bit);
1138 if (first_set >= bitmap_size)
1139 return false;
1140
1141 *found_start = folio_start + (first_set << fs_info->sectorsize_bits);
1142 first_zero = find_next_zero_bit(delalloc_bitmap, bitmap_size, first_set);
1143 *found_len = (first_zero - first_set) << fs_info->sectorsize_bits;
1144 return true;
1145}
1146
1147/*
1148 * Do all of the delayed allocation setup.
1149 *
1150 * Return >0 if all the dirty blocks are submitted async (compression) or inlined.
1151 * The @folio should no longer be touched (treat it as already unlocked).
1152 *
1153 * Return 0 if there is still dirty block that needs to be submitted through
1154 * extent_writepage_io().
1155 * bio_ctrl->submit_bitmap will indicate which blocks of the folio should be
1156 * submitted, and @folio is still kept locked.
1157 *
1158 * Return <0 if there is any error hit.
1159 * Any allocated ordered extent range covering this folio will be marked
1160 * finished (IOERR), and @folio is still kept locked.
1161 */
1162static noinline_for_stack int writepage_delalloc(struct btrfs_inode *inode,
1163 struct folio *folio,
1164 struct btrfs_bio_ctrl *bio_ctrl)
1165{
1166 struct btrfs_fs_info *fs_info = inode_to_fs_info(&inode->vfs_inode);
1167 struct writeback_control *wbc = bio_ctrl->wbc;
1168 const bool is_subpage = btrfs_is_subpage(fs_info, folio->mapping);
1169 const u64 page_start = folio_pos(folio);
1170 const u64 page_end = page_start + folio_size(folio) - 1;
1171 unsigned long delalloc_bitmap = 0;
1172 /*
1173 * Save the last found delalloc end. As the delalloc end can go beyond
1174 * page boundary, thus we cannot rely on subpage bitmap to locate the
1175 * last delalloc end.
1176 */
1177 u64 last_delalloc_end = 0;
1178 /*
1179 * The range end (exclusive) of the last successfully finished delalloc
1180 * range.
1181 * Any range covered by ordered extent must either be manually marked
1182 * finished (error handling), or has IO submitted (and finish the
1183 * ordered extent normally).
1184 *
1185 * This records the end of ordered extent cleanup if we hit an error.
1186 */
1187 u64 last_finished_delalloc_end = page_start;
1188 u64 delalloc_start = page_start;
1189 u64 delalloc_end = page_end;
1190 u64 delalloc_to_write = 0;
1191 int ret = 0;
1192 int bit;
1193
1194 /* Save the dirty bitmap as our submission bitmap will be a subset of it. */
1195 if (btrfs_is_subpage(fs_info, inode->vfs_inode.i_mapping)) {
1196 ASSERT(fs_info->sectors_per_page > 1);
1197 btrfs_get_subpage_dirty_bitmap(fs_info, folio, &bio_ctrl->submit_bitmap);
1198 } else {
1199 bio_ctrl->submit_bitmap = 1;
1200 }
1201
1202 for_each_set_bit(bit, &bio_ctrl->submit_bitmap, fs_info->sectors_per_page) {
1203 u64 start = page_start + (bit << fs_info->sectorsize_bits);
1204
1205 btrfs_folio_set_lock(fs_info, folio, start, fs_info->sectorsize);
1206 }
1207
1208 /* Lock all (subpage) delalloc ranges inside the folio first. */
1209 while (delalloc_start < page_end) {
1210 delalloc_end = page_end;
1211 if (!find_lock_delalloc_range(&inode->vfs_inode, folio,
1212 &delalloc_start, &delalloc_end)) {
1213 delalloc_start = delalloc_end + 1;
1214 continue;
1215 }
1216 set_delalloc_bitmap(folio, &delalloc_bitmap, delalloc_start,
1217 min(delalloc_end, page_end) + 1 - delalloc_start);
1218 last_delalloc_end = delalloc_end;
1219 delalloc_start = delalloc_end + 1;
1220 }
1221 delalloc_start = page_start;
1222
1223 if (!last_delalloc_end)
1224 goto out;
1225
1226 /* Run the delalloc ranges for the above locked ranges. */
1227 while (delalloc_start < page_end) {
1228 u64 found_start;
1229 u32 found_len;
1230 bool found;
1231
1232 if (!is_subpage) {
1233 /*
1234 * For non-subpage case, the found delalloc range must
1235 * cover this folio and there must be only one locked
1236 * delalloc range.
1237 */
1238 found_start = page_start;
1239 found_len = last_delalloc_end + 1 - found_start;
1240 found = true;
1241 } else {
1242 found = find_next_delalloc_bitmap(folio, &delalloc_bitmap,
1243 delalloc_start, &found_start, &found_len);
1244 }
1245 if (!found)
1246 break;
1247 /*
1248 * The subpage range covers the last sector, the delalloc range may
1249 * end beyond the folio boundary, use the saved delalloc_end
1250 * instead.
1251 */
1252 if (found_start + found_len >= page_end)
1253 found_len = last_delalloc_end + 1 - found_start;
1254
1255 if (ret >= 0) {
1256 /*
1257 * Some delalloc range may be created by previous folios.
1258 * Thus we still need to clean up this range during error
1259 * handling.
1260 */
1261 last_finished_delalloc_end = found_start;
1262 /* No errors hit so far, run the current delalloc range. */
1263 ret = btrfs_run_delalloc_range(inode, folio,
1264 found_start,
1265 found_start + found_len - 1,
1266 wbc);
1267 if (ret >= 0)
1268 last_finished_delalloc_end = found_start + found_len;
1269 } else {
1270 /*
1271 * We've hit an error during previous delalloc range,
1272 * have to cleanup the remaining locked ranges.
1273 */
1274 unlock_extent(&inode->io_tree, found_start,
1275 found_start + found_len - 1, NULL);
1276 __unlock_for_delalloc(&inode->vfs_inode, folio,
1277 found_start,
1278 found_start + found_len - 1);
1279 }
1280
1281 /*
1282 * We have some ranges that's going to be submitted asynchronously
1283 * (compression or inline). These range have their own control
1284 * on when to unlock the pages. We should not touch them
1285 * anymore, so clear the range from the submission bitmap.
1286 */
1287 if (ret > 0) {
1288 unsigned int start_bit = (found_start - page_start) >>
1289 fs_info->sectorsize_bits;
1290 unsigned int end_bit = (min(page_end + 1, found_start + found_len) -
1291 page_start) >> fs_info->sectorsize_bits;
1292 bitmap_clear(&bio_ctrl->submit_bitmap, start_bit, end_bit - start_bit);
1293 }
1294 /*
1295 * Above btrfs_run_delalloc_range() may have unlocked the folio,
1296 * thus for the last range, we cannot touch the folio anymore.
1297 */
1298 if (found_start + found_len >= last_delalloc_end + 1)
1299 break;
1300
1301 delalloc_start = found_start + found_len;
1302 }
1303 /*
1304 * It's possible we had some ordered extents created before we hit
1305 * an error, cleanup non-async successfully created delalloc ranges.
1306 */
1307 if (unlikely(ret < 0)) {
1308 unsigned int bitmap_size = min(
1309 (last_finished_delalloc_end - page_start) >>
1310 fs_info->sectorsize_bits,
1311 fs_info->sectors_per_page);
1312
1313 for_each_set_bit(bit, &bio_ctrl->submit_bitmap, bitmap_size)
1314 btrfs_mark_ordered_io_finished(inode, folio,
1315 page_start + (bit << fs_info->sectorsize_bits),
1316 fs_info->sectorsize, false);
1317 return ret;
1318 }
1319out:
1320 if (last_delalloc_end)
1321 delalloc_end = last_delalloc_end;
1322 else
1323 delalloc_end = page_end;
1324 /*
1325 * delalloc_end is already one less than the total length, so
1326 * we don't subtract one from PAGE_SIZE
1327 */
1328 delalloc_to_write +=
1329 DIV_ROUND_UP(delalloc_end + 1 - page_start, PAGE_SIZE);
1330
1331 /*
1332 * If all ranges are submitted asynchronously, we just need to account
1333 * for them here.
1334 */
1335 if (bitmap_empty(&bio_ctrl->submit_bitmap, fs_info->sectors_per_page)) {
1336 wbc->nr_to_write -= delalloc_to_write;
1337 return 1;
1338 }
1339
1340 if (wbc->nr_to_write < delalloc_to_write) {
1341 int thresh = 8192;
1342
1343 if (delalloc_to_write < thresh * 2)
1344 thresh = delalloc_to_write;
1345 wbc->nr_to_write = min_t(u64, delalloc_to_write,
1346 thresh);
1347 }
1348
1349 return 0;
1350}
1351
1352/*
1353 * Return 0 if we have submitted or queued the sector for submission.
1354 * Return <0 for critical errors.
1355 *
1356 * Caller should make sure filepos < i_size and handle filepos >= i_size case.
1357 */
1358static int submit_one_sector(struct btrfs_inode *inode,
1359 struct folio *folio,
1360 u64 filepos, struct btrfs_bio_ctrl *bio_ctrl,
1361 loff_t i_size)
1362{
1363 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1364 struct extent_map *em;
1365 u64 block_start;
1366 u64 disk_bytenr;
1367 u64 extent_offset;
1368 u64 em_end;
1369 const u32 sectorsize = fs_info->sectorsize;
1370
1371 ASSERT(IS_ALIGNED(filepos, sectorsize));
1372
1373 /* @filepos >= i_size case should be handled by the caller. */
1374 ASSERT(filepos < i_size);
1375
1376 em = btrfs_get_extent(inode, NULL, filepos, sectorsize);
1377 if (IS_ERR(em))
1378 return PTR_ERR_OR_ZERO(em);
1379
1380 extent_offset = filepos - em->start;
1381 em_end = extent_map_end(em);
1382 ASSERT(filepos <= em_end);
1383 ASSERT(IS_ALIGNED(em->start, sectorsize));
1384 ASSERT(IS_ALIGNED(em->len, sectorsize));
1385
1386 block_start = extent_map_block_start(em);
1387 disk_bytenr = extent_map_block_start(em) + extent_offset;
1388
1389 ASSERT(!extent_map_is_compressed(em));
1390 ASSERT(block_start != EXTENT_MAP_HOLE);
1391 ASSERT(block_start != EXTENT_MAP_INLINE);
1392
1393 free_extent_map(em);
1394 em = NULL;
1395
1396 /*
1397 * Although the PageDirty bit is cleared before entering this
1398 * function, subpage dirty bit is not cleared.
1399 * So clear subpage dirty bit here so next time we won't submit
1400 * a folio for a range already written to disk.
1401 */
1402 btrfs_folio_clear_dirty(fs_info, folio, filepos, sectorsize);
1403 btrfs_folio_set_writeback(fs_info, folio, filepos, sectorsize);
1404 /*
1405 * Above call should set the whole folio with writeback flag, even
1406 * just for a single subpage sector.
1407 * As long as the folio is properly locked and the range is correct,
1408 * we should always get the folio with writeback flag.
1409 */
1410 ASSERT(folio_test_writeback(folio));
1411
1412 submit_extent_folio(bio_ctrl, disk_bytenr, folio,
1413 sectorsize, filepos - folio_pos(folio));
1414 return 0;
1415}
1416
1417/*
1418 * Helper for extent_writepage(). This calls the writepage start hooks,
1419 * and does the loop to map the page into extents and bios.
1420 *
1421 * We return 1 if the IO is started and the page is unlocked,
1422 * 0 if all went well (page still locked)
1423 * < 0 if there were errors (page still locked)
1424 */
1425static noinline_for_stack int extent_writepage_io(struct btrfs_inode *inode,
1426 struct folio *folio,
1427 u64 start, u32 len,
1428 struct btrfs_bio_ctrl *bio_ctrl,
1429 loff_t i_size)
1430{
1431 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1432 unsigned long range_bitmap = 0;
1433 bool submitted_io = false;
1434 bool error = false;
1435 const u64 folio_start = folio_pos(folio);
1436 u64 cur;
1437 int bit;
1438 int ret = 0;
1439
1440 ASSERT(start >= folio_start &&
1441 start + len <= folio_start + folio_size(folio));
1442
1443 ret = btrfs_writepage_cow_fixup(folio);
1444 if (ret) {
1445 /* Fixup worker will requeue */
1446 folio_redirty_for_writepage(bio_ctrl->wbc, folio);
1447 folio_unlock(folio);
1448 return 1;
1449 }
1450
1451 for (cur = start; cur < start + len; cur += fs_info->sectorsize)
1452 set_bit((cur - folio_start) >> fs_info->sectorsize_bits, &range_bitmap);
1453 bitmap_and(&bio_ctrl->submit_bitmap, &bio_ctrl->submit_bitmap, &range_bitmap,
1454 fs_info->sectors_per_page);
1455
1456 bio_ctrl->end_io_func = end_bbio_data_write;
1457
1458 for_each_set_bit(bit, &bio_ctrl->submit_bitmap, fs_info->sectors_per_page) {
1459 cur = folio_pos(folio) + (bit << fs_info->sectorsize_bits);
1460
1461 if (cur >= i_size) {
1462 btrfs_mark_ordered_io_finished(inode, folio, cur,
1463 start + len - cur, true);
1464 /*
1465 * This range is beyond i_size, thus we don't need to
1466 * bother writing back.
1467 * But we still need to clear the dirty subpage bit, or
1468 * the next time the folio gets dirtied, we will try to
1469 * writeback the sectors with subpage dirty bits,
1470 * causing writeback without ordered extent.
1471 */
1472 btrfs_folio_clear_dirty(fs_info, folio, cur,
1473 start + len - cur);
1474 break;
1475 }
1476 ret = submit_one_sector(inode, folio, cur, bio_ctrl, i_size);
1477 if (unlikely(ret < 0)) {
1478 /*
1479 * bio_ctrl may contain a bio crossing several folios.
1480 * Submit it immediately so that the bio has a chance
1481 * to finish normally, other than marked as error.
1482 */
1483 submit_one_bio(bio_ctrl);
1484 /*
1485 * Failed to grab the extent map which should be very rare.
1486 * Since there is no bio submitted to finish the ordered
1487 * extent, we have to manually finish this sector.
1488 */
1489 btrfs_mark_ordered_io_finished(inode, folio, cur,
1490 fs_info->sectorsize, false);
1491 error = true;
1492 continue;
1493 }
1494 submitted_io = true;
1495 }
1496
1497 /*
1498 * If we didn't submitted any sector (>= i_size), folio dirty get
1499 * cleared but PAGECACHE_TAG_DIRTY is not cleared (only cleared
1500 * by folio_start_writeback() if the folio is not dirty).
1501 *
1502 * Here we set writeback and clear for the range. If the full folio
1503 * is no longer dirty then we clear the PAGECACHE_TAG_DIRTY tag.
1504 *
1505 * If we hit any error, the corresponding sector will still be dirty
1506 * thus no need to clear PAGECACHE_TAG_DIRTY.
1507 */
1508 if (!submitted_io && !error) {
1509 btrfs_folio_set_writeback(fs_info, folio, start, len);
1510 btrfs_folio_clear_writeback(fs_info, folio, start, len);
1511 }
1512 return ret;
1513}
1514
1515/*
1516 * the writepage semantics are similar to regular writepage. extent
1517 * records are inserted to lock ranges in the tree, and as dirty areas
1518 * are found, they are marked writeback. Then the lock bits are removed
1519 * and the end_io handler clears the writeback ranges
1520 *
1521 * Return 0 if everything goes well.
1522 * Return <0 for error.
1523 */
1524static int extent_writepage(struct folio *folio, struct btrfs_bio_ctrl *bio_ctrl)
1525{
1526 struct btrfs_inode *inode = BTRFS_I(folio->mapping->host);
1527 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1528 int ret;
1529 size_t pg_offset;
1530 loff_t i_size = i_size_read(&inode->vfs_inode);
1531 unsigned long end_index = i_size >> PAGE_SHIFT;
1532
1533 trace_extent_writepage(folio, &inode->vfs_inode, bio_ctrl->wbc);
1534
1535 WARN_ON(!folio_test_locked(folio));
1536
1537 pg_offset = offset_in_folio(folio, i_size);
1538 if (folio->index > end_index ||
1539 (folio->index == end_index && !pg_offset)) {
1540 folio_invalidate(folio, 0, folio_size(folio));
1541 folio_unlock(folio);
1542 return 0;
1543 }
1544
1545 if (folio->index == end_index)
1546 folio_zero_range(folio, pg_offset, folio_size(folio) - pg_offset);
1547
1548 /*
1549 * Default to unlock the whole folio.
1550 * The proper bitmap can only be initialized until writepage_delalloc().
1551 */
1552 bio_ctrl->submit_bitmap = (unsigned long)-1;
1553 ret = set_folio_extent_mapped(folio);
1554 if (ret < 0)
1555 goto done;
1556
1557 ret = writepage_delalloc(inode, folio, bio_ctrl);
1558 if (ret == 1)
1559 return 0;
1560 if (ret)
1561 goto done;
1562
1563 ret = extent_writepage_io(inode, folio, folio_pos(folio),
1564 PAGE_SIZE, bio_ctrl, i_size);
1565 if (ret == 1)
1566 return 0;
1567
1568 bio_ctrl->wbc->nr_to_write--;
1569
1570done:
1571 if (ret < 0)
1572 mapping_set_error(folio->mapping, ret);
1573 /*
1574 * Only unlock ranges that are submitted. As there can be some async
1575 * submitted ranges inside the folio.
1576 */
1577 btrfs_folio_end_lock_bitmap(fs_info, folio, bio_ctrl->submit_bitmap);
1578 ASSERT(ret <= 0);
1579 return ret;
1580}
1581
1582void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
1583{
1584 wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
1585 TASK_UNINTERRUPTIBLE);
1586}
1587
1588/*
1589 * Lock extent buffer status and pages for writeback.
1590 *
1591 * Return %false if the extent buffer doesn't need to be submitted (e.g. the
1592 * extent buffer is not dirty)
1593 * Return %true is the extent buffer is submitted to bio.
1594 */
1595static noinline_for_stack bool lock_extent_buffer_for_io(struct extent_buffer *eb,
1596 struct writeback_control *wbc)
1597{
1598 struct btrfs_fs_info *fs_info = eb->fs_info;
1599 bool ret = false;
1600
1601 btrfs_tree_lock(eb);
1602 while (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
1603 btrfs_tree_unlock(eb);
1604 if (wbc->sync_mode != WB_SYNC_ALL)
1605 return false;
1606 wait_on_extent_buffer_writeback(eb);
1607 btrfs_tree_lock(eb);
1608 }
1609
1610 /*
1611 * We need to do this to prevent races in people who check if the eb is
1612 * under IO since we can end up having no IO bits set for a short period
1613 * of time.
1614 */
1615 spin_lock(&eb->refs_lock);
1616 if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
1617 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
1618 spin_unlock(&eb->refs_lock);
1619 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
1620 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
1621 -eb->len,
1622 fs_info->dirty_metadata_batch);
1623 ret = true;
1624 } else {
1625 spin_unlock(&eb->refs_lock);
1626 }
1627 btrfs_tree_unlock(eb);
1628 return ret;
1629}
1630
1631static void set_btree_ioerr(struct extent_buffer *eb)
1632{
1633 struct btrfs_fs_info *fs_info = eb->fs_info;
1634
1635 set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
1636
1637 /*
1638 * A read may stumble upon this buffer later, make sure that it gets an
1639 * error and knows there was an error.
1640 */
1641 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
1642
1643 /*
1644 * We need to set the mapping with the io error as well because a write
1645 * error will flip the file system readonly, and then syncfs() will
1646 * return a 0 because we are readonly if we don't modify the err seq for
1647 * the superblock.
1648 */
1649 mapping_set_error(eb->fs_info->btree_inode->i_mapping, -EIO);
1650
1651 /*
1652 * If writeback for a btree extent that doesn't belong to a log tree
1653 * failed, increment the counter transaction->eb_write_errors.
1654 * We do this because while the transaction is running and before it's
1655 * committing (when we call filemap_fdata[write|wait]_range against
1656 * the btree inode), we might have
1657 * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
1658 * returns an error or an error happens during writeback, when we're
1659 * committing the transaction we wouldn't know about it, since the pages
1660 * can be no longer dirty nor marked anymore for writeback (if a
1661 * subsequent modification to the extent buffer didn't happen before the
1662 * transaction commit), which makes filemap_fdata[write|wait]_range not
1663 * able to find the pages which contain errors at transaction
1664 * commit time. So if this happens we must abort the transaction,
1665 * otherwise we commit a super block with btree roots that point to
1666 * btree nodes/leafs whose content on disk is invalid - either garbage
1667 * or the content of some node/leaf from a past generation that got
1668 * cowed or deleted and is no longer valid.
1669 *
1670 * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
1671 * not be enough - we need to distinguish between log tree extents vs
1672 * non-log tree extents, and the next filemap_fdatawait_range() call
1673 * will catch and clear such errors in the mapping - and that call might
1674 * be from a log sync and not from a transaction commit. Also, checking
1675 * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
1676 * not done and would not be reliable - the eb might have been released
1677 * from memory and reading it back again means that flag would not be
1678 * set (since it's a runtime flag, not persisted on disk).
1679 *
1680 * Using the flags below in the btree inode also makes us achieve the
1681 * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
1682 * writeback for all dirty pages and before filemap_fdatawait_range()
1683 * is called, the writeback for all dirty pages had already finished
1684 * with errors - because we were not using AS_EIO/AS_ENOSPC,
1685 * filemap_fdatawait_range() would return success, as it could not know
1686 * that writeback errors happened (the pages were no longer tagged for
1687 * writeback).
1688 */
1689 switch (eb->log_index) {
1690 case -1:
1691 set_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags);
1692 break;
1693 case 0:
1694 set_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
1695 break;
1696 case 1:
1697 set_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
1698 break;
1699 default:
1700 BUG(); /* unexpected, logic error */
1701 }
1702}
1703
1704/*
1705 * The endio specific version which won't touch any unsafe spinlock in endio
1706 * context.
1707 */
1708static struct extent_buffer *find_extent_buffer_nolock(
1709 const struct btrfs_fs_info *fs_info, u64 start)
1710{
1711 struct extent_buffer *eb;
1712
1713 rcu_read_lock();
1714 eb = radix_tree_lookup(&fs_info->buffer_radix,
1715 start >> fs_info->sectorsize_bits);
1716 if (eb && atomic_inc_not_zero(&eb->refs)) {
1717 rcu_read_unlock();
1718 return eb;
1719 }
1720 rcu_read_unlock();
1721 return NULL;
1722}
1723
1724static void end_bbio_meta_write(struct btrfs_bio *bbio)
1725{
1726 struct extent_buffer *eb = bbio->private;
1727 struct btrfs_fs_info *fs_info = eb->fs_info;
1728 bool uptodate = !bbio->bio.bi_status;
1729 struct folio_iter fi;
1730 u32 bio_offset = 0;
1731
1732 if (!uptodate)
1733 set_btree_ioerr(eb);
1734
1735 bio_for_each_folio_all(fi, &bbio->bio) {
1736 u64 start = eb->start + bio_offset;
1737 struct folio *folio = fi.folio;
1738 u32 len = fi.length;
1739
1740 btrfs_folio_clear_writeback(fs_info, folio, start, len);
1741 bio_offset += len;
1742 }
1743
1744 clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
1745 smp_mb__after_atomic();
1746 wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
1747
1748 bio_put(&bbio->bio);
1749}
1750
1751static void prepare_eb_write(struct extent_buffer *eb)
1752{
1753 u32 nritems;
1754 unsigned long start;
1755 unsigned long end;
1756
1757 clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
1758
1759 /* Set btree blocks beyond nritems with 0 to avoid stale content */
1760 nritems = btrfs_header_nritems(eb);
1761 if (btrfs_header_level(eb) > 0) {
1762 end = btrfs_node_key_ptr_offset(eb, nritems);
1763 memzero_extent_buffer(eb, end, eb->len - end);
1764 } else {
1765 /*
1766 * Leaf:
1767 * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
1768 */
1769 start = btrfs_item_nr_offset(eb, nritems);
1770 end = btrfs_item_nr_offset(eb, 0);
1771 if (nritems == 0)
1772 end += BTRFS_LEAF_DATA_SIZE(eb->fs_info);
1773 else
1774 end += btrfs_item_offset(eb, nritems - 1);
1775 memzero_extent_buffer(eb, start, end - start);
1776 }
1777}
1778
1779static noinline_for_stack void write_one_eb(struct extent_buffer *eb,
1780 struct writeback_control *wbc)
1781{
1782 struct btrfs_fs_info *fs_info = eb->fs_info;
1783 struct btrfs_bio *bbio;
1784
1785 prepare_eb_write(eb);
1786
1787 bbio = btrfs_bio_alloc(INLINE_EXTENT_BUFFER_PAGES,
1788 REQ_OP_WRITE | REQ_META | wbc_to_write_flags(wbc),
1789 eb->fs_info, end_bbio_meta_write, eb);
1790 bbio->bio.bi_iter.bi_sector = eb->start >> SECTOR_SHIFT;
1791 bio_set_dev(&bbio->bio, fs_info->fs_devices->latest_dev->bdev);
1792 wbc_init_bio(wbc, &bbio->bio);
1793 bbio->inode = BTRFS_I(eb->fs_info->btree_inode);
1794 bbio->file_offset = eb->start;
1795 if (fs_info->nodesize < PAGE_SIZE) {
1796 struct folio *folio = eb->folios[0];
1797 bool ret;
1798
1799 folio_lock(folio);
1800 btrfs_subpage_set_writeback(fs_info, folio, eb->start, eb->len);
1801 if (btrfs_subpage_clear_and_test_dirty(fs_info, folio, eb->start,
1802 eb->len)) {
1803 folio_clear_dirty_for_io(folio);
1804 wbc->nr_to_write--;
1805 }
1806 ret = bio_add_folio(&bbio->bio, folio, eb->len,
1807 eb->start - folio_pos(folio));
1808 ASSERT(ret);
1809 wbc_account_cgroup_owner(wbc, folio, eb->len);
1810 folio_unlock(folio);
1811 } else {
1812 int num_folios = num_extent_folios(eb);
1813
1814 for (int i = 0; i < num_folios; i++) {
1815 struct folio *folio = eb->folios[i];
1816 bool ret;
1817
1818 folio_lock(folio);
1819 folio_clear_dirty_for_io(folio);
1820 folio_start_writeback(folio);
1821 ret = bio_add_folio(&bbio->bio, folio, eb->folio_size, 0);
1822 ASSERT(ret);
1823 wbc_account_cgroup_owner(wbc, folio, eb->folio_size);
1824 wbc->nr_to_write -= folio_nr_pages(folio);
1825 folio_unlock(folio);
1826 }
1827 }
1828 btrfs_submit_bbio(bbio, 0);
1829}
1830
1831/*
1832 * Submit one subpage btree page.
1833 *
1834 * The main difference to submit_eb_page() is:
1835 * - Page locking
1836 * For subpage, we don't rely on page locking at all.
1837 *
1838 * - Flush write bio
1839 * We only flush bio if we may be unable to fit current extent buffers into
1840 * current bio.
1841 *
1842 * Return >=0 for the number of submitted extent buffers.
1843 * Return <0 for fatal error.
1844 */
1845static int submit_eb_subpage(struct folio *folio, struct writeback_control *wbc)
1846{
1847 struct btrfs_fs_info *fs_info = folio_to_fs_info(folio);
1848 int submitted = 0;
1849 u64 folio_start = folio_pos(folio);
1850 int bit_start = 0;
1851 int sectors_per_node = fs_info->nodesize >> fs_info->sectorsize_bits;
1852
1853 /* Lock and write each dirty extent buffers in the range */
1854 while (bit_start < fs_info->sectors_per_page) {
1855 struct btrfs_subpage *subpage = folio_get_private(folio);
1856 struct extent_buffer *eb;
1857 unsigned long flags;
1858 u64 start;
1859
1860 /*
1861 * Take private lock to ensure the subpage won't be detached
1862 * in the meantime.
1863 */
1864 spin_lock(&folio->mapping->i_private_lock);
1865 if (!folio_test_private(folio)) {
1866 spin_unlock(&folio->mapping->i_private_lock);
1867 break;
1868 }
1869 spin_lock_irqsave(&subpage->lock, flags);
1870 if (!test_bit(bit_start + btrfs_bitmap_nr_dirty * fs_info->sectors_per_page,
1871 subpage->bitmaps)) {
1872 spin_unlock_irqrestore(&subpage->lock, flags);
1873 spin_unlock(&folio->mapping->i_private_lock);
1874 bit_start++;
1875 continue;
1876 }
1877
1878 start = folio_start + bit_start * fs_info->sectorsize;
1879 bit_start += sectors_per_node;
1880
1881 /*
1882 * Here we just want to grab the eb without touching extra
1883 * spin locks, so call find_extent_buffer_nolock().
1884 */
1885 eb = find_extent_buffer_nolock(fs_info, start);
1886 spin_unlock_irqrestore(&subpage->lock, flags);
1887 spin_unlock(&folio->mapping->i_private_lock);
1888
1889 /*
1890 * The eb has already reached 0 refs thus find_extent_buffer()
1891 * doesn't return it. We don't need to write back such eb
1892 * anyway.
1893 */
1894 if (!eb)
1895 continue;
1896
1897 if (lock_extent_buffer_for_io(eb, wbc)) {
1898 write_one_eb(eb, wbc);
1899 submitted++;
1900 }
1901 free_extent_buffer(eb);
1902 }
1903 return submitted;
1904}
1905
1906/*
1907 * Submit all page(s) of one extent buffer.
1908 *
1909 * @page: the page of one extent buffer
1910 * @eb_context: to determine if we need to submit this page, if current page
1911 * belongs to this eb, we don't need to submit
1912 *
1913 * The caller should pass each page in their bytenr order, and here we use
1914 * @eb_context to determine if we have submitted pages of one extent buffer.
1915 *
1916 * If we have, we just skip until we hit a new page that doesn't belong to
1917 * current @eb_context.
1918 *
1919 * If not, we submit all the page(s) of the extent buffer.
1920 *
1921 * Return >0 if we have submitted the extent buffer successfully.
1922 * Return 0 if we don't need to submit the page, as it's already submitted by
1923 * previous call.
1924 * Return <0 for fatal error.
1925 */
1926static int submit_eb_page(struct folio *folio, struct btrfs_eb_write_context *ctx)
1927{
1928 struct writeback_control *wbc = ctx->wbc;
1929 struct address_space *mapping = folio->mapping;
1930 struct extent_buffer *eb;
1931 int ret;
1932
1933 if (!folio_test_private(folio))
1934 return 0;
1935
1936 if (folio_to_fs_info(folio)->nodesize < PAGE_SIZE)
1937 return submit_eb_subpage(folio, wbc);
1938
1939 spin_lock(&mapping->i_private_lock);
1940 if (!folio_test_private(folio)) {
1941 spin_unlock(&mapping->i_private_lock);
1942 return 0;
1943 }
1944
1945 eb = folio_get_private(folio);
1946
1947 /*
1948 * Shouldn't happen and normally this would be a BUG_ON but no point
1949 * crashing the machine for something we can survive anyway.
1950 */
1951 if (WARN_ON(!eb)) {
1952 spin_unlock(&mapping->i_private_lock);
1953 return 0;
1954 }
1955
1956 if (eb == ctx->eb) {
1957 spin_unlock(&mapping->i_private_lock);
1958 return 0;
1959 }
1960 ret = atomic_inc_not_zero(&eb->refs);
1961 spin_unlock(&mapping->i_private_lock);
1962 if (!ret)
1963 return 0;
1964
1965 ctx->eb = eb;
1966
1967 ret = btrfs_check_meta_write_pointer(eb->fs_info, ctx);
1968 if (ret) {
1969 if (ret == -EBUSY)
1970 ret = 0;
1971 free_extent_buffer(eb);
1972 return ret;
1973 }
1974
1975 if (!lock_extent_buffer_for_io(eb, wbc)) {
1976 free_extent_buffer(eb);
1977 return 0;
1978 }
1979 /* Implies write in zoned mode. */
1980 if (ctx->zoned_bg) {
1981 /* Mark the last eb in the block group. */
1982 btrfs_schedule_zone_finish_bg(ctx->zoned_bg, eb);
1983 ctx->zoned_bg->meta_write_pointer += eb->len;
1984 }
1985 write_one_eb(eb, wbc);
1986 free_extent_buffer(eb);
1987 return 1;
1988}
1989
1990int btree_write_cache_pages(struct address_space *mapping,
1991 struct writeback_control *wbc)
1992{
1993 struct btrfs_eb_write_context ctx = { .wbc = wbc };
1994 struct btrfs_fs_info *fs_info = inode_to_fs_info(mapping->host);
1995 int ret = 0;
1996 int done = 0;
1997 int nr_to_write_done = 0;
1998 struct folio_batch fbatch;
1999 unsigned int nr_folios;
2000 pgoff_t index;
2001 pgoff_t end; /* Inclusive */
2002 int scanned = 0;
2003 xa_mark_t tag;
2004
2005 folio_batch_init(&fbatch);
2006 if (wbc->range_cyclic) {
2007 index = mapping->writeback_index; /* Start from prev offset */
2008 end = -1;
2009 /*
2010 * Start from the beginning does not need to cycle over the
2011 * range, mark it as scanned.
2012 */
2013 scanned = (index == 0);
2014 } else {
2015 index = wbc->range_start >> PAGE_SHIFT;
2016 end = wbc->range_end >> PAGE_SHIFT;
2017 scanned = 1;
2018 }
2019 if (wbc->sync_mode == WB_SYNC_ALL)
2020 tag = PAGECACHE_TAG_TOWRITE;
2021 else
2022 tag = PAGECACHE_TAG_DIRTY;
2023 btrfs_zoned_meta_io_lock(fs_info);
2024retry:
2025 if (wbc->sync_mode == WB_SYNC_ALL)
2026 tag_pages_for_writeback(mapping, index, end);
2027 while (!done && !nr_to_write_done && (index <= end) &&
2028 (nr_folios = filemap_get_folios_tag(mapping, &index, end,
2029 tag, &fbatch))) {
2030 unsigned i;
2031
2032 for (i = 0; i < nr_folios; i++) {
2033 struct folio *folio = fbatch.folios[i];
2034
2035 ret = submit_eb_page(folio, &ctx);
2036 if (ret == 0)
2037 continue;
2038 if (ret < 0) {
2039 done = 1;
2040 break;
2041 }
2042
2043 /*
2044 * the filesystem may choose to bump up nr_to_write.
2045 * We have to make sure to honor the new nr_to_write
2046 * at any time
2047 */
2048 nr_to_write_done = wbc->nr_to_write <= 0;
2049 }
2050 folio_batch_release(&fbatch);
2051 cond_resched();
2052 }
2053 if (!scanned && !done) {
2054 /*
2055 * We hit the last page and there is more work to be done: wrap
2056 * back to the start of the file
2057 */
2058 scanned = 1;
2059 index = 0;
2060 goto retry;
2061 }
2062 /*
2063 * If something went wrong, don't allow any metadata write bio to be
2064 * submitted.
2065 *
2066 * This would prevent use-after-free if we had dirty pages not
2067 * cleaned up, which can still happen by fuzzed images.
2068 *
2069 * - Bad extent tree
2070 * Allowing existing tree block to be allocated for other trees.
2071 *
2072 * - Log tree operations
2073 * Exiting tree blocks get allocated to log tree, bumps its
2074 * generation, then get cleaned in tree re-balance.
2075 * Such tree block will not be written back, since it's clean,
2076 * thus no WRITTEN flag set.
2077 * And after log writes back, this tree block is not traced by
2078 * any dirty extent_io_tree.
2079 *
2080 * - Offending tree block gets re-dirtied from its original owner
2081 * Since it has bumped generation, no WRITTEN flag, it can be
2082 * reused without COWing. This tree block will not be traced
2083 * by btrfs_transaction::dirty_pages.
2084 *
2085 * Now such dirty tree block will not be cleaned by any dirty
2086 * extent io tree. Thus we don't want to submit such wild eb
2087 * if the fs already has error.
2088 *
2089 * We can get ret > 0 from submit_extent_folio() indicating how many ebs
2090 * were submitted. Reset it to 0 to avoid false alerts for the caller.
2091 */
2092 if (ret > 0)
2093 ret = 0;
2094 if (!ret && BTRFS_FS_ERROR(fs_info))
2095 ret = -EROFS;
2096
2097 if (ctx.zoned_bg)
2098 btrfs_put_block_group(ctx.zoned_bg);
2099 btrfs_zoned_meta_io_unlock(fs_info);
2100 return ret;
2101}
2102
2103/*
2104 * Walk the list of dirty pages of the given address space and write all of them.
2105 *
2106 * @mapping: address space structure to write
2107 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2108 * @bio_ctrl: holds context for the write, namely the bio
2109 *
2110 * If a page is already under I/O, write_cache_pages() skips it, even
2111 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2112 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2113 * and msync() need to guarantee that all the data which was dirty at the time
2114 * the call was made get new I/O started against them. If wbc->sync_mode is
2115 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2116 * existing IO to complete.
2117 */
2118static int extent_write_cache_pages(struct address_space *mapping,
2119 struct btrfs_bio_ctrl *bio_ctrl)
2120{
2121 struct writeback_control *wbc = bio_ctrl->wbc;
2122 struct inode *inode = mapping->host;
2123 int ret = 0;
2124 int done = 0;
2125 int nr_to_write_done = 0;
2126 struct folio_batch fbatch;
2127 unsigned int nr_folios;
2128 pgoff_t index;
2129 pgoff_t end; /* Inclusive */
2130 pgoff_t done_index;
2131 int range_whole = 0;
2132 int scanned = 0;
2133 xa_mark_t tag;
2134
2135 /*
2136 * We have to hold onto the inode so that ordered extents can do their
2137 * work when the IO finishes. The alternative to this is failing to add
2138 * an ordered extent if the igrab() fails there and that is a huge pain
2139 * to deal with, so instead just hold onto the inode throughout the
2140 * writepages operation. If it fails here we are freeing up the inode
2141 * anyway and we'd rather not waste our time writing out stuff that is
2142 * going to be truncated anyway.
2143 */
2144 if (!igrab(inode))
2145 return 0;
2146
2147 folio_batch_init(&fbatch);
2148 if (wbc->range_cyclic) {
2149 index = mapping->writeback_index; /* Start from prev offset */
2150 end = -1;
2151 /*
2152 * Start from the beginning does not need to cycle over the
2153 * range, mark it as scanned.
2154 */
2155 scanned = (index == 0);
2156 } else {
2157 index = wbc->range_start >> PAGE_SHIFT;
2158 end = wbc->range_end >> PAGE_SHIFT;
2159 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2160 range_whole = 1;
2161 scanned = 1;
2162 }
2163
2164 /*
2165 * We do the tagged writepage as long as the snapshot flush bit is set
2166 * and we are the first one who do the filemap_flush() on this inode.
2167 *
2168 * The nr_to_write == LONG_MAX is needed to make sure other flushers do
2169 * not race in and drop the bit.
2170 */
2171 if (range_whole && wbc->nr_to_write == LONG_MAX &&
2172 test_and_clear_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
2173 &BTRFS_I(inode)->runtime_flags))
2174 wbc->tagged_writepages = 1;
2175
2176 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2177 tag = PAGECACHE_TAG_TOWRITE;
2178 else
2179 tag = PAGECACHE_TAG_DIRTY;
2180retry:
2181 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2182 tag_pages_for_writeback(mapping, index, end);
2183 done_index = index;
2184 while (!done && !nr_to_write_done && (index <= end) &&
2185 (nr_folios = filemap_get_folios_tag(mapping, &index,
2186 end, tag, &fbatch))) {
2187 unsigned i;
2188
2189 for (i = 0; i < nr_folios; i++) {
2190 struct folio *folio = fbatch.folios[i];
2191
2192 done_index = folio_next_index(folio);
2193 /*
2194 * At this point we hold neither the i_pages lock nor
2195 * the page lock: the page may be truncated or
2196 * invalidated (changing page->mapping to NULL),
2197 * or even swizzled back from swapper_space to
2198 * tmpfs file mapping
2199 */
2200 if (!folio_trylock(folio)) {
2201 submit_write_bio(bio_ctrl, 0);
2202 folio_lock(folio);
2203 }
2204
2205 if (unlikely(folio->mapping != mapping)) {
2206 folio_unlock(folio);
2207 continue;
2208 }
2209
2210 if (!folio_test_dirty(folio)) {
2211 /* Someone wrote it for us. */
2212 folio_unlock(folio);
2213 continue;
2214 }
2215
2216 /*
2217 * For subpage case, compression can lead to mixed
2218 * writeback and dirty flags, e.g:
2219 * 0 32K 64K 96K 128K
2220 * | |//////||/////| |//|
2221 *
2222 * In above case, [32K, 96K) is asynchronously submitted
2223 * for compression, and [124K, 128K) needs to be written back.
2224 *
2225 * If we didn't wait wrtiteback for page 64K, [128K, 128K)
2226 * won't be submitted as the page still has writeback flag
2227 * and will be skipped in the next check.
2228 *
2229 * This mixed writeback and dirty case is only possible for
2230 * subpage case.
2231 *
2232 * TODO: Remove this check after migrating compression to
2233 * regular submission.
2234 */
2235 if (wbc->sync_mode != WB_SYNC_NONE ||
2236 btrfs_is_subpage(inode_to_fs_info(inode), mapping)) {
2237 if (folio_test_writeback(folio))
2238 submit_write_bio(bio_ctrl, 0);
2239 folio_wait_writeback(folio);
2240 }
2241
2242 if (folio_test_writeback(folio) ||
2243 !folio_clear_dirty_for_io(folio)) {
2244 folio_unlock(folio);
2245 continue;
2246 }
2247
2248 ret = extent_writepage(folio, bio_ctrl);
2249 if (ret < 0) {
2250 done = 1;
2251 break;
2252 }
2253
2254 /*
2255 * The filesystem may choose to bump up nr_to_write.
2256 * We have to make sure to honor the new nr_to_write
2257 * at any time.
2258 */
2259 nr_to_write_done = (wbc->sync_mode == WB_SYNC_NONE &&
2260 wbc->nr_to_write <= 0);
2261 }
2262 folio_batch_release(&fbatch);
2263 cond_resched();
2264 }
2265 if (!scanned && !done) {
2266 /*
2267 * We hit the last page and there is more work to be done: wrap
2268 * back to the start of the file
2269 */
2270 scanned = 1;
2271 index = 0;
2272
2273 /*
2274 * If we're looping we could run into a page that is locked by a
2275 * writer and that writer could be waiting on writeback for a
2276 * page in our current bio, and thus deadlock, so flush the
2277 * write bio here.
2278 */
2279 submit_write_bio(bio_ctrl, 0);
2280 goto retry;
2281 }
2282
2283 if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
2284 mapping->writeback_index = done_index;
2285
2286 btrfs_add_delayed_iput(BTRFS_I(inode));
2287 return ret;
2288}
2289
2290/*
2291 * Submit the pages in the range to bio for call sites which delalloc range has
2292 * already been ran (aka, ordered extent inserted) and all pages are still
2293 * locked.
2294 */
2295void extent_write_locked_range(struct inode *inode, const struct folio *locked_folio,
2296 u64 start, u64 end, struct writeback_control *wbc,
2297 bool pages_dirty)
2298{
2299 bool found_error = false;
2300 int ret = 0;
2301 struct address_space *mapping = inode->i_mapping;
2302 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
2303 const u32 sectorsize = fs_info->sectorsize;
2304 loff_t i_size = i_size_read(inode);
2305 u64 cur = start;
2306 struct btrfs_bio_ctrl bio_ctrl = {
2307 .wbc = wbc,
2308 .opf = REQ_OP_WRITE | wbc_to_write_flags(wbc),
2309 };
2310
2311 if (wbc->no_cgroup_owner)
2312 bio_ctrl.opf |= REQ_BTRFS_CGROUP_PUNT;
2313
2314 ASSERT(IS_ALIGNED(start, sectorsize) && IS_ALIGNED(end + 1, sectorsize));
2315
2316 while (cur <= end) {
2317 u64 cur_end = min(round_down(cur, PAGE_SIZE) + PAGE_SIZE - 1, end);
2318 u32 cur_len = cur_end + 1 - cur;
2319 struct folio *folio;
2320
2321 folio = filemap_get_folio(mapping, cur >> PAGE_SHIFT);
2322
2323 /*
2324 * This shouldn't happen, the pages are pinned and locked, this
2325 * code is just in case, but shouldn't actually be run.
2326 */
2327 if (IS_ERR(folio)) {
2328 btrfs_mark_ordered_io_finished(BTRFS_I(inode), NULL,
2329 cur, cur_len, false);
2330 mapping_set_error(mapping, PTR_ERR(folio));
2331 cur = cur_end + 1;
2332 continue;
2333 }
2334
2335 ASSERT(folio_test_locked(folio));
2336 if (pages_dirty && folio != locked_folio)
2337 ASSERT(folio_test_dirty(folio));
2338
2339 /*
2340 * Set the submission bitmap to submit all sectors.
2341 * extent_writepage_io() will do the truncation correctly.
2342 */
2343 bio_ctrl.submit_bitmap = (unsigned long)-1;
2344 ret = extent_writepage_io(BTRFS_I(inode), folio, cur, cur_len,
2345 &bio_ctrl, i_size);
2346 if (ret == 1)
2347 goto next_page;
2348
2349 if (ret)
2350 mapping_set_error(mapping, ret);
2351 btrfs_folio_end_lock(fs_info, folio, cur, cur_len);
2352 if (ret < 0)
2353 found_error = true;
2354next_page:
2355 folio_put(folio);
2356 cur = cur_end + 1;
2357 }
2358
2359 submit_write_bio(&bio_ctrl, found_error ? ret : 0);
2360}
2361
2362int btrfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
2363{
2364 struct inode *inode = mapping->host;
2365 int ret = 0;
2366 struct btrfs_bio_ctrl bio_ctrl = {
2367 .wbc = wbc,
2368 .opf = REQ_OP_WRITE | wbc_to_write_flags(wbc),
2369 };
2370
2371 /*
2372 * Allow only a single thread to do the reloc work in zoned mode to
2373 * protect the write pointer updates.
2374 */
2375 btrfs_zoned_data_reloc_lock(BTRFS_I(inode));
2376 ret = extent_write_cache_pages(mapping, &bio_ctrl);
2377 submit_write_bio(&bio_ctrl, ret);
2378 btrfs_zoned_data_reloc_unlock(BTRFS_I(inode));
2379 return ret;
2380}
2381
2382void btrfs_readahead(struct readahead_control *rac)
2383{
2384 struct btrfs_bio_ctrl bio_ctrl = { .opf = REQ_OP_READ | REQ_RAHEAD };
2385 struct folio *folio;
2386 struct btrfs_inode *inode = BTRFS_I(rac->mapping->host);
2387 const u64 start = readahead_pos(rac);
2388 const u64 end = start + readahead_length(rac) - 1;
2389 struct extent_state *cached_state = NULL;
2390 struct extent_map *em_cached = NULL;
2391 u64 prev_em_start = (u64)-1;
2392
2393 btrfs_lock_and_flush_ordered_range(inode, start, end, &cached_state);
2394
2395 while ((folio = readahead_folio(rac)) != NULL)
2396 btrfs_do_readpage(folio, &em_cached, &bio_ctrl, &prev_em_start);
2397
2398 unlock_extent(&inode->io_tree, start, end, &cached_state);
2399
2400 if (em_cached)
2401 free_extent_map(em_cached);
2402 submit_one_bio(&bio_ctrl);
2403}
2404
2405/*
2406 * basic invalidate_folio code, this waits on any locked or writeback
2407 * ranges corresponding to the folio, and then deletes any extent state
2408 * records from the tree
2409 */
2410int extent_invalidate_folio(struct extent_io_tree *tree,
2411 struct folio *folio, size_t offset)
2412{
2413 struct extent_state *cached_state = NULL;
2414 u64 start = folio_pos(folio);
2415 u64 end = start + folio_size(folio) - 1;
2416 size_t blocksize = folio_to_fs_info(folio)->sectorsize;
2417
2418 /* This function is only called for the btree inode */
2419 ASSERT(tree->owner == IO_TREE_BTREE_INODE_IO);
2420
2421 start += ALIGN(offset, blocksize);
2422 if (start > end)
2423 return 0;
2424
2425 lock_extent(tree, start, end, &cached_state);
2426 folio_wait_writeback(folio);
2427
2428 /*
2429 * Currently for btree io tree, only EXTENT_LOCKED is utilized,
2430 * so here we only need to unlock the extent range to free any
2431 * existing extent state.
2432 */
2433 unlock_extent(tree, start, end, &cached_state);
2434 return 0;
2435}
2436
2437/*
2438 * a helper for release_folio, this tests for areas of the page that
2439 * are locked or under IO and drops the related state bits if it is safe
2440 * to drop the page.
2441 */
2442static bool try_release_extent_state(struct extent_io_tree *tree,
2443 struct folio *folio)
2444{
2445 u64 start = folio_pos(folio);
2446 u64 end = start + PAGE_SIZE - 1;
2447 bool ret;
2448
2449 if (test_range_bit_exists(tree, start, end, EXTENT_LOCKED)) {
2450 ret = false;
2451 } else {
2452 u32 clear_bits = ~(EXTENT_LOCKED | EXTENT_NODATASUM |
2453 EXTENT_DELALLOC_NEW | EXTENT_CTLBITS |
2454 EXTENT_QGROUP_RESERVED);
2455 int ret2;
2456
2457 /*
2458 * At this point we can safely clear everything except the
2459 * locked bit, the nodatasum bit and the delalloc new bit.
2460 * The delalloc new bit will be cleared by ordered extent
2461 * completion.
2462 */
2463 ret2 = __clear_extent_bit(tree, start, end, clear_bits, NULL, NULL);
2464
2465 /* if clear_extent_bit failed for enomem reasons,
2466 * we can't allow the release to continue.
2467 */
2468 if (ret2 < 0)
2469 ret = false;
2470 else
2471 ret = true;
2472 }
2473 return ret;
2474}
2475
2476/*
2477 * a helper for release_folio. As long as there are no locked extents
2478 * in the range corresponding to the page, both state records and extent
2479 * map records are removed
2480 */
2481bool try_release_extent_mapping(struct folio *folio, gfp_t mask)
2482{
2483 u64 start = folio_pos(folio);
2484 u64 end = start + PAGE_SIZE - 1;
2485 struct btrfs_inode *inode = folio_to_inode(folio);
2486 struct extent_io_tree *io_tree = &inode->io_tree;
2487
2488 while (start <= end) {
2489 const u64 cur_gen = btrfs_get_fs_generation(inode->root->fs_info);
2490 const u64 len = end - start + 1;
2491 struct extent_map_tree *extent_tree = &inode->extent_tree;
2492 struct extent_map *em;
2493
2494 write_lock(&extent_tree->lock);
2495 em = lookup_extent_mapping(extent_tree, start, len);
2496 if (!em) {
2497 write_unlock(&extent_tree->lock);
2498 break;
2499 }
2500 if ((em->flags & EXTENT_FLAG_PINNED) || em->start != start) {
2501 write_unlock(&extent_tree->lock);
2502 free_extent_map(em);
2503 break;
2504 }
2505 if (test_range_bit_exists(io_tree, em->start,
2506 extent_map_end(em) - 1, EXTENT_LOCKED))
2507 goto next;
2508 /*
2509 * If it's not in the list of modified extents, used by a fast
2510 * fsync, we can remove it. If it's being logged we can safely
2511 * remove it since fsync took an extra reference on the em.
2512 */
2513 if (list_empty(&em->list) || (em->flags & EXTENT_FLAG_LOGGING))
2514 goto remove_em;
2515 /*
2516 * If it's in the list of modified extents, remove it only if
2517 * its generation is older then the current one, in which case
2518 * we don't need it for a fast fsync. Otherwise don't remove it,
2519 * we could be racing with an ongoing fast fsync that could miss
2520 * the new extent.
2521 */
2522 if (em->generation >= cur_gen)
2523 goto next;
2524remove_em:
2525 /*
2526 * We only remove extent maps that are not in the list of
2527 * modified extents or that are in the list but with a
2528 * generation lower then the current generation, so there is no
2529 * need to set the full fsync flag on the inode (it hurts the
2530 * fsync performance for workloads with a data size that exceeds
2531 * or is close to the system's memory).
2532 */
2533 remove_extent_mapping(inode, em);
2534 /* Once for the inode's extent map tree. */
2535 free_extent_map(em);
2536next:
2537 start = extent_map_end(em);
2538 write_unlock(&extent_tree->lock);
2539
2540 /* Once for us, for the lookup_extent_mapping() reference. */
2541 free_extent_map(em);
2542
2543 if (need_resched()) {
2544 /*
2545 * If we need to resched but we can't block just exit
2546 * and leave any remaining extent maps.
2547 */
2548 if (!gfpflags_allow_blocking(mask))
2549 break;
2550
2551 cond_resched();
2552 }
2553 }
2554 return try_release_extent_state(io_tree, folio);
2555}
2556
2557static void __free_extent_buffer(struct extent_buffer *eb)
2558{
2559 kmem_cache_free(extent_buffer_cache, eb);
2560}
2561
2562static int extent_buffer_under_io(const struct extent_buffer *eb)
2563{
2564 return (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
2565 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
2566}
2567
2568static bool folio_range_has_eb(struct folio *folio)
2569{
2570 struct btrfs_subpage *subpage;
2571
2572 lockdep_assert_held(&folio->mapping->i_private_lock);
2573
2574 if (folio_test_private(folio)) {
2575 subpage = folio_get_private(folio);
2576 if (atomic_read(&subpage->eb_refs))
2577 return true;
2578 }
2579 return false;
2580}
2581
2582static void detach_extent_buffer_folio(const struct extent_buffer *eb, struct folio *folio)
2583{
2584 struct btrfs_fs_info *fs_info = eb->fs_info;
2585 const bool mapped = !test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
2586
2587 /*
2588 * For mapped eb, we're going to change the folio private, which should
2589 * be done under the i_private_lock.
2590 */
2591 if (mapped)
2592 spin_lock(&folio->mapping->i_private_lock);
2593
2594 if (!folio_test_private(folio)) {
2595 if (mapped)
2596 spin_unlock(&folio->mapping->i_private_lock);
2597 return;
2598 }
2599
2600 if (fs_info->nodesize >= PAGE_SIZE) {
2601 /*
2602 * We do this since we'll remove the pages after we've
2603 * removed the eb from the radix tree, so we could race
2604 * and have this page now attached to the new eb. So
2605 * only clear folio if it's still connected to
2606 * this eb.
2607 */
2608 if (folio_test_private(folio) && folio_get_private(folio) == eb) {
2609 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
2610 BUG_ON(folio_test_dirty(folio));
2611 BUG_ON(folio_test_writeback(folio));
2612 /* We need to make sure we haven't be attached to a new eb. */
2613 folio_detach_private(folio);
2614 }
2615 if (mapped)
2616 spin_unlock(&folio->mapping->i_private_lock);
2617 return;
2618 }
2619
2620 /*
2621 * For subpage, we can have dummy eb with folio private attached. In
2622 * this case, we can directly detach the private as such folio is only
2623 * attached to one dummy eb, no sharing.
2624 */
2625 if (!mapped) {
2626 btrfs_detach_subpage(fs_info, folio);
2627 return;
2628 }
2629
2630 btrfs_folio_dec_eb_refs(fs_info, folio);
2631
2632 /*
2633 * We can only detach the folio private if there are no other ebs in the
2634 * page range and no unfinished IO.
2635 */
2636 if (!folio_range_has_eb(folio))
2637 btrfs_detach_subpage(fs_info, folio);
2638
2639 spin_unlock(&folio->mapping->i_private_lock);
2640}
2641
2642/* Release all pages attached to the extent buffer */
2643static void btrfs_release_extent_buffer_pages(const struct extent_buffer *eb)
2644{
2645 ASSERT(!extent_buffer_under_io(eb));
2646
2647 for (int i = 0; i < INLINE_EXTENT_BUFFER_PAGES; i++) {
2648 struct folio *folio = eb->folios[i];
2649
2650 if (!folio)
2651 continue;
2652
2653 detach_extent_buffer_folio(eb, folio);
2654
2655 /* One for when we allocated the folio. */
2656 folio_put(folio);
2657 }
2658}
2659
2660/*
2661 * Helper for releasing the extent buffer.
2662 */
2663static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
2664{
2665 btrfs_release_extent_buffer_pages(eb);
2666 btrfs_leak_debug_del_eb(eb);
2667 __free_extent_buffer(eb);
2668}
2669
2670static struct extent_buffer *
2671__alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
2672 unsigned long len)
2673{
2674 struct extent_buffer *eb = NULL;
2675
2676 eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
2677 eb->start = start;
2678 eb->len = len;
2679 eb->fs_info = fs_info;
2680 init_rwsem(&eb->lock);
2681
2682 btrfs_leak_debug_add_eb(eb);
2683
2684 spin_lock_init(&eb->refs_lock);
2685 atomic_set(&eb->refs, 1);
2686
2687 ASSERT(len <= BTRFS_MAX_METADATA_BLOCKSIZE);
2688
2689 return eb;
2690}
2691
2692struct extent_buffer *btrfs_clone_extent_buffer(const struct extent_buffer *src)
2693{
2694 struct extent_buffer *new;
2695 int num_folios = num_extent_folios(src);
2696 int ret;
2697
2698 new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
2699 if (new == NULL)
2700 return NULL;
2701
2702 /*
2703 * Set UNMAPPED before calling btrfs_release_extent_buffer(), as
2704 * btrfs_release_extent_buffer() have different behavior for
2705 * UNMAPPED subpage extent buffer.
2706 */
2707 set_bit(EXTENT_BUFFER_UNMAPPED, &new->bflags);
2708
2709 ret = alloc_eb_folio_array(new, false);
2710 if (ret) {
2711 btrfs_release_extent_buffer(new);
2712 return NULL;
2713 }
2714
2715 for (int i = 0; i < num_folios; i++) {
2716 struct folio *folio = new->folios[i];
2717
2718 ret = attach_extent_buffer_folio(new, folio, NULL);
2719 if (ret < 0) {
2720 btrfs_release_extent_buffer(new);
2721 return NULL;
2722 }
2723 WARN_ON(folio_test_dirty(folio));
2724 }
2725 copy_extent_buffer_full(new, src);
2726 set_extent_buffer_uptodate(new);
2727
2728 return new;
2729}
2730
2731struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
2732 u64 start, unsigned long len)
2733{
2734 struct extent_buffer *eb;
2735 int num_folios = 0;
2736 int ret;
2737
2738 eb = __alloc_extent_buffer(fs_info, start, len);
2739 if (!eb)
2740 return NULL;
2741
2742 ret = alloc_eb_folio_array(eb, false);
2743 if (ret)
2744 goto err;
2745
2746 num_folios = num_extent_folios(eb);
2747 for (int i = 0; i < num_folios; i++) {
2748 ret = attach_extent_buffer_folio(eb, eb->folios[i], NULL);
2749 if (ret < 0)
2750 goto err;
2751 }
2752
2753 set_extent_buffer_uptodate(eb);
2754 btrfs_set_header_nritems(eb, 0);
2755 set_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
2756
2757 return eb;
2758err:
2759 for (int i = 0; i < num_folios; i++) {
2760 if (eb->folios[i]) {
2761 detach_extent_buffer_folio(eb, eb->folios[i]);
2762 folio_put(eb->folios[i]);
2763 }
2764 }
2765 __free_extent_buffer(eb);
2766 return NULL;
2767}
2768
2769struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
2770 u64 start)
2771{
2772 return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
2773}
2774
2775static void check_buffer_tree_ref(struct extent_buffer *eb)
2776{
2777 int refs;
2778 /*
2779 * The TREE_REF bit is first set when the extent_buffer is added
2780 * to the radix tree. It is also reset, if unset, when a new reference
2781 * is created by find_extent_buffer.
2782 *
2783 * It is only cleared in two cases: freeing the last non-tree
2784 * reference to the extent_buffer when its STALE bit is set or
2785 * calling release_folio when the tree reference is the only reference.
2786 *
2787 * In both cases, care is taken to ensure that the extent_buffer's
2788 * pages are not under io. However, release_folio can be concurrently
2789 * called with creating new references, which is prone to race
2790 * conditions between the calls to check_buffer_tree_ref in those
2791 * codepaths and clearing TREE_REF in try_release_extent_buffer.
2792 *
2793 * The actual lifetime of the extent_buffer in the radix tree is
2794 * adequately protected by the refcount, but the TREE_REF bit and
2795 * its corresponding reference are not. To protect against this
2796 * class of races, we call check_buffer_tree_ref from the codepaths
2797 * which trigger io. Note that once io is initiated, TREE_REF can no
2798 * longer be cleared, so that is the moment at which any such race is
2799 * best fixed.
2800 */
2801 refs = atomic_read(&eb->refs);
2802 if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
2803 return;
2804
2805 spin_lock(&eb->refs_lock);
2806 if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
2807 atomic_inc(&eb->refs);
2808 spin_unlock(&eb->refs_lock);
2809}
2810
2811static void mark_extent_buffer_accessed(struct extent_buffer *eb)
2812{
2813 int num_folios= num_extent_folios(eb);
2814
2815 check_buffer_tree_ref(eb);
2816
2817 for (int i = 0; i < num_folios; i++)
2818 folio_mark_accessed(eb->folios[i]);
2819}
2820
2821struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
2822 u64 start)
2823{
2824 struct extent_buffer *eb;
2825
2826 eb = find_extent_buffer_nolock(fs_info, start);
2827 if (!eb)
2828 return NULL;
2829 /*
2830 * Lock our eb's refs_lock to avoid races with free_extent_buffer().
2831 * When we get our eb it might be flagged with EXTENT_BUFFER_STALE and
2832 * another task running free_extent_buffer() might have seen that flag
2833 * set, eb->refs == 2, that the buffer isn't under IO (dirty and
2834 * writeback flags not set) and it's still in the tree (flag
2835 * EXTENT_BUFFER_TREE_REF set), therefore being in the process of
2836 * decrementing the extent buffer's reference count twice. So here we
2837 * could race and increment the eb's reference count, clear its stale
2838 * flag, mark it as dirty and drop our reference before the other task
2839 * finishes executing free_extent_buffer, which would later result in
2840 * an attempt to free an extent buffer that is dirty.
2841 */
2842 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
2843 spin_lock(&eb->refs_lock);
2844 spin_unlock(&eb->refs_lock);
2845 }
2846 mark_extent_buffer_accessed(eb);
2847 return eb;
2848}
2849
2850#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
2851struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
2852 u64 start)
2853{
2854 struct extent_buffer *eb, *exists = NULL;
2855 int ret;
2856
2857 eb = find_extent_buffer(fs_info, start);
2858 if (eb)
2859 return eb;
2860 eb = alloc_dummy_extent_buffer(fs_info, start);
2861 if (!eb)
2862 return ERR_PTR(-ENOMEM);
2863 eb->fs_info = fs_info;
2864again:
2865 ret = radix_tree_preload(GFP_NOFS);
2866 if (ret) {
2867 exists = ERR_PTR(ret);
2868 goto free_eb;
2869 }
2870 spin_lock(&fs_info->buffer_lock);
2871 ret = radix_tree_insert(&fs_info->buffer_radix,
2872 start >> fs_info->sectorsize_bits, eb);
2873 spin_unlock(&fs_info->buffer_lock);
2874 radix_tree_preload_end();
2875 if (ret == -EEXIST) {
2876 exists = find_extent_buffer(fs_info, start);
2877 if (exists)
2878 goto free_eb;
2879 else
2880 goto again;
2881 }
2882 check_buffer_tree_ref(eb);
2883 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
2884
2885 return eb;
2886free_eb:
2887 btrfs_release_extent_buffer(eb);
2888 return exists;
2889}
2890#endif
2891
2892static struct extent_buffer *grab_extent_buffer(
2893 struct btrfs_fs_info *fs_info, struct page *page)
2894{
2895 struct folio *folio = page_folio(page);
2896 struct extent_buffer *exists;
2897
2898 lockdep_assert_held(&page->mapping->i_private_lock);
2899
2900 /*
2901 * For subpage case, we completely rely on radix tree to ensure we
2902 * don't try to insert two ebs for the same bytenr. So here we always
2903 * return NULL and just continue.
2904 */
2905 if (fs_info->nodesize < PAGE_SIZE)
2906 return NULL;
2907
2908 /* Page not yet attached to an extent buffer */
2909 if (!folio_test_private(folio))
2910 return NULL;
2911
2912 /*
2913 * We could have already allocated an eb for this page and attached one
2914 * so lets see if we can get a ref on the existing eb, and if we can we
2915 * know it's good and we can just return that one, else we know we can
2916 * just overwrite folio private.
2917 */
2918 exists = folio_get_private(folio);
2919 if (atomic_inc_not_zero(&exists->refs))
2920 return exists;
2921
2922 WARN_ON(PageDirty(page));
2923 folio_detach_private(folio);
2924 return NULL;
2925}
2926
2927static int check_eb_alignment(struct btrfs_fs_info *fs_info, u64 start)
2928{
2929 if (!IS_ALIGNED(start, fs_info->sectorsize)) {
2930 btrfs_err(fs_info, "bad tree block start %llu", start);
2931 return -EINVAL;
2932 }
2933
2934 if (fs_info->nodesize < PAGE_SIZE &&
2935 offset_in_page(start) + fs_info->nodesize > PAGE_SIZE) {
2936 btrfs_err(fs_info,
2937 "tree block crosses page boundary, start %llu nodesize %u",
2938 start, fs_info->nodesize);
2939 return -EINVAL;
2940 }
2941 if (fs_info->nodesize >= PAGE_SIZE &&
2942 !PAGE_ALIGNED(start)) {
2943 btrfs_err(fs_info,
2944 "tree block is not page aligned, start %llu nodesize %u",
2945 start, fs_info->nodesize);
2946 return -EINVAL;
2947 }
2948 if (!IS_ALIGNED(start, fs_info->nodesize) &&
2949 !test_and_set_bit(BTRFS_FS_UNALIGNED_TREE_BLOCK, &fs_info->flags)) {
2950 btrfs_warn(fs_info,
2951"tree block not nodesize aligned, start %llu nodesize %u, can be resolved by a full metadata balance",
2952 start, fs_info->nodesize);
2953 }
2954 return 0;
2955}
2956
2957
2958/*
2959 * Return 0 if eb->folios[i] is attached to btree inode successfully.
2960 * Return >0 if there is already another extent buffer for the range,
2961 * and @found_eb_ret would be updated.
2962 * Return -EAGAIN if the filemap has an existing folio but with different size
2963 * than @eb.
2964 * The caller needs to free the existing folios and retry using the same order.
2965 */
2966static int attach_eb_folio_to_filemap(struct extent_buffer *eb, int i,
2967 struct btrfs_subpage *prealloc,
2968 struct extent_buffer **found_eb_ret)
2969{
2970
2971 struct btrfs_fs_info *fs_info = eb->fs_info;
2972 struct address_space *mapping = fs_info->btree_inode->i_mapping;
2973 const unsigned long index = eb->start >> PAGE_SHIFT;
2974 struct folio *existing_folio = NULL;
2975 int ret;
2976
2977 ASSERT(found_eb_ret);
2978
2979 /* Caller should ensure the folio exists. */
2980 ASSERT(eb->folios[i]);
2981
2982retry:
2983 ret = filemap_add_folio(mapping, eb->folios[i], index + i,
2984 GFP_NOFS | __GFP_NOFAIL);
2985 if (!ret)
2986 goto finish;
2987
2988 existing_folio = filemap_lock_folio(mapping, index + i);
2989 /* The page cache only exists for a very short time, just retry. */
2990 if (IS_ERR(existing_folio)) {
2991 existing_folio = NULL;
2992 goto retry;
2993 }
2994
2995 /* For now, we should only have single-page folios for btree inode. */
2996 ASSERT(folio_nr_pages(existing_folio) == 1);
2997
2998 if (folio_size(existing_folio) != eb->folio_size) {
2999 folio_unlock(existing_folio);
3000 folio_put(existing_folio);
3001 return -EAGAIN;
3002 }
3003
3004finish:
3005 spin_lock(&mapping->i_private_lock);
3006 if (existing_folio && fs_info->nodesize < PAGE_SIZE) {
3007 /* We're going to reuse the existing page, can drop our folio now. */
3008 __free_page(folio_page(eb->folios[i], 0));
3009 eb->folios[i] = existing_folio;
3010 } else if (existing_folio) {
3011 struct extent_buffer *existing_eb;
3012
3013 existing_eb = grab_extent_buffer(fs_info,
3014 folio_page(existing_folio, 0));
3015 if (existing_eb) {
3016 /* The extent buffer still exists, we can use it directly. */
3017 *found_eb_ret = existing_eb;
3018 spin_unlock(&mapping->i_private_lock);
3019 folio_unlock(existing_folio);
3020 folio_put(existing_folio);
3021 return 1;
3022 }
3023 /* The extent buffer no longer exists, we can reuse the folio. */
3024 __free_page(folio_page(eb->folios[i], 0));
3025 eb->folios[i] = existing_folio;
3026 }
3027 eb->folio_size = folio_size(eb->folios[i]);
3028 eb->folio_shift = folio_shift(eb->folios[i]);
3029 /* Should not fail, as we have preallocated the memory. */
3030 ret = attach_extent_buffer_folio(eb, eb->folios[i], prealloc);
3031 ASSERT(!ret);
3032 /*
3033 * To inform we have an extra eb under allocation, so that
3034 * detach_extent_buffer_page() won't release the folio private when the
3035 * eb hasn't been inserted into radix tree yet.
3036 *
3037 * The ref will be decreased when the eb releases the page, in
3038 * detach_extent_buffer_page(). Thus needs no special handling in the
3039 * error path.
3040 */
3041 btrfs_folio_inc_eb_refs(fs_info, eb->folios[i]);
3042 spin_unlock(&mapping->i_private_lock);
3043 return 0;
3044}
3045
3046struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
3047 u64 start, u64 owner_root, int level)
3048{
3049 unsigned long len = fs_info->nodesize;
3050 int num_folios;
3051 int attached = 0;
3052 struct extent_buffer *eb;
3053 struct extent_buffer *existing_eb = NULL;
3054 struct btrfs_subpage *prealloc = NULL;
3055 u64 lockdep_owner = owner_root;
3056 bool page_contig = true;
3057 int uptodate = 1;
3058 int ret;
3059
3060 if (check_eb_alignment(fs_info, start))
3061 return ERR_PTR(-EINVAL);
3062
3063#if BITS_PER_LONG == 32
3064 if (start >= MAX_LFS_FILESIZE) {
3065 btrfs_err_rl(fs_info,
3066 "extent buffer %llu is beyond 32bit page cache limit", start);
3067 btrfs_err_32bit_limit(fs_info);
3068 return ERR_PTR(-EOVERFLOW);
3069 }
3070 if (start >= BTRFS_32BIT_EARLY_WARN_THRESHOLD)
3071 btrfs_warn_32bit_limit(fs_info);
3072#endif
3073
3074 eb = find_extent_buffer(fs_info, start);
3075 if (eb)
3076 return eb;
3077
3078 eb = __alloc_extent_buffer(fs_info, start, len);
3079 if (!eb)
3080 return ERR_PTR(-ENOMEM);
3081
3082 /*
3083 * The reloc trees are just snapshots, so we need them to appear to be
3084 * just like any other fs tree WRT lockdep.
3085 */
3086 if (lockdep_owner == BTRFS_TREE_RELOC_OBJECTID)
3087 lockdep_owner = BTRFS_FS_TREE_OBJECTID;
3088
3089 btrfs_set_buffer_lockdep_class(lockdep_owner, eb, level);
3090
3091 /*
3092 * Preallocate folio private for subpage case, so that we won't
3093 * allocate memory with i_private_lock nor page lock hold.
3094 *
3095 * The memory will be freed by attach_extent_buffer_page() or freed
3096 * manually if we exit earlier.
3097 */
3098 if (fs_info->nodesize < PAGE_SIZE) {
3099 prealloc = btrfs_alloc_subpage(fs_info, BTRFS_SUBPAGE_METADATA);
3100 if (IS_ERR(prealloc)) {
3101 ret = PTR_ERR(prealloc);
3102 goto out;
3103 }
3104 }
3105
3106reallocate:
3107 /* Allocate all pages first. */
3108 ret = alloc_eb_folio_array(eb, true);
3109 if (ret < 0) {
3110 btrfs_free_subpage(prealloc);
3111 goto out;
3112 }
3113
3114 num_folios = num_extent_folios(eb);
3115 /* Attach all pages to the filemap. */
3116 for (int i = 0; i < num_folios; i++) {
3117 struct folio *folio;
3118
3119 ret = attach_eb_folio_to_filemap(eb, i, prealloc, &existing_eb);
3120 if (ret > 0) {
3121 ASSERT(existing_eb);
3122 goto out;
3123 }
3124
3125 /*
3126 * TODO: Special handling for a corner case where the order of
3127 * folios mismatch between the new eb and filemap.
3128 *
3129 * This happens when:
3130 *
3131 * - the new eb is using higher order folio
3132 *
3133 * - the filemap is still using 0-order folios for the range
3134 * This can happen at the previous eb allocation, and we don't
3135 * have higher order folio for the call.
3136 *
3137 * - the existing eb has already been freed
3138 *
3139 * In this case, we have to free the existing folios first, and
3140 * re-allocate using the same order.
3141 * Thankfully this is not going to happen yet, as we're still
3142 * using 0-order folios.
3143 */
3144 if (unlikely(ret == -EAGAIN)) {
3145 ASSERT(0);
3146 goto reallocate;
3147 }
3148 attached++;
3149
3150 /*
3151 * Only after attach_eb_folio_to_filemap(), eb->folios[] is
3152 * reliable, as we may choose to reuse the existing page cache
3153 * and free the allocated page.
3154 */
3155 folio = eb->folios[i];
3156 WARN_ON(btrfs_folio_test_dirty(fs_info, folio, eb->start, eb->len));
3157
3158 /*
3159 * Check if the current page is physically contiguous with previous eb
3160 * page.
3161 * At this stage, either we allocated a large folio, thus @i
3162 * would only be 0, or we fall back to per-page allocation.
3163 */
3164 if (i && folio_page(eb->folios[i - 1], 0) + 1 != folio_page(folio, 0))
3165 page_contig = false;
3166
3167 if (!btrfs_folio_test_uptodate(fs_info, folio, eb->start, eb->len))
3168 uptodate = 0;
3169
3170 /*
3171 * We can't unlock the pages just yet since the extent buffer
3172 * hasn't been properly inserted in the radix tree, this
3173 * opens a race with btree_release_folio which can free a page
3174 * while we are still filling in all pages for the buffer and
3175 * we could crash.
3176 */
3177 }
3178 if (uptodate)
3179 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3180 /* All pages are physically contiguous, can skip cross page handling. */
3181 if (page_contig)
3182 eb->addr = folio_address(eb->folios[0]) + offset_in_page(eb->start);
3183again:
3184 ret = radix_tree_preload(GFP_NOFS);
3185 if (ret)
3186 goto out;
3187
3188 spin_lock(&fs_info->buffer_lock);
3189 ret = radix_tree_insert(&fs_info->buffer_radix,
3190 start >> fs_info->sectorsize_bits, eb);
3191 spin_unlock(&fs_info->buffer_lock);
3192 radix_tree_preload_end();
3193 if (ret == -EEXIST) {
3194 ret = 0;
3195 existing_eb = find_extent_buffer(fs_info, start);
3196 if (existing_eb)
3197 goto out;
3198 else
3199 goto again;
3200 }
3201 /* add one reference for the tree */
3202 check_buffer_tree_ref(eb);
3203 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
3204
3205 /*
3206 * Now it's safe to unlock the pages because any calls to
3207 * btree_release_folio will correctly detect that a page belongs to a
3208 * live buffer and won't free them prematurely.
3209 */
3210 for (int i = 0; i < num_folios; i++)
3211 unlock_page(folio_page(eb->folios[i], 0));
3212 return eb;
3213
3214out:
3215 WARN_ON(!atomic_dec_and_test(&eb->refs));
3216
3217 /*
3218 * Any attached folios need to be detached before we unlock them. This
3219 * is because when we're inserting our new folios into the mapping, and
3220 * then attaching our eb to that folio. If we fail to insert our folio
3221 * we'll lookup the folio for that index, and grab that EB. We do not
3222 * want that to grab this eb, as we're getting ready to free it. So we
3223 * have to detach it first and then unlock it.
3224 *
3225 * We have to drop our reference and NULL it out here because in the
3226 * subpage case detaching does a btrfs_folio_dec_eb_refs() for our eb.
3227 * Below when we call btrfs_release_extent_buffer() we will call
3228 * detach_extent_buffer_folio() on our remaining pages in the !subpage
3229 * case. If we left eb->folios[i] populated in the subpage case we'd
3230 * double put our reference and be super sad.
3231 */
3232 for (int i = 0; i < attached; i++) {
3233 ASSERT(eb->folios[i]);
3234 detach_extent_buffer_folio(eb, eb->folios[i]);
3235 unlock_page(folio_page(eb->folios[i], 0));
3236 folio_put(eb->folios[i]);
3237 eb->folios[i] = NULL;
3238 }
3239 /*
3240 * Now all pages of that extent buffer is unmapped, set UNMAPPED flag,
3241 * so it can be cleaned up without utilizing page->mapping.
3242 */
3243 set_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
3244
3245 btrfs_release_extent_buffer(eb);
3246 if (ret < 0)
3247 return ERR_PTR(ret);
3248 ASSERT(existing_eb);
3249 return existing_eb;
3250}
3251
3252static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
3253{
3254 struct extent_buffer *eb =
3255 container_of(head, struct extent_buffer, rcu_head);
3256
3257 __free_extent_buffer(eb);
3258}
3259
3260static int release_extent_buffer(struct extent_buffer *eb)
3261 __releases(&eb->refs_lock)
3262{
3263 lockdep_assert_held(&eb->refs_lock);
3264
3265 WARN_ON(atomic_read(&eb->refs) == 0);
3266 if (atomic_dec_and_test(&eb->refs)) {
3267 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
3268 struct btrfs_fs_info *fs_info = eb->fs_info;
3269
3270 spin_unlock(&eb->refs_lock);
3271
3272 spin_lock(&fs_info->buffer_lock);
3273 radix_tree_delete(&fs_info->buffer_radix,
3274 eb->start >> fs_info->sectorsize_bits);
3275 spin_unlock(&fs_info->buffer_lock);
3276 } else {
3277 spin_unlock(&eb->refs_lock);
3278 }
3279
3280 btrfs_leak_debug_del_eb(eb);
3281 /* Should be safe to release our pages at this point */
3282 btrfs_release_extent_buffer_pages(eb);
3283#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3284 if (unlikely(test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags))) {
3285 __free_extent_buffer(eb);
3286 return 1;
3287 }
3288#endif
3289 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
3290 return 1;
3291 }
3292 spin_unlock(&eb->refs_lock);
3293
3294 return 0;
3295}
3296
3297void free_extent_buffer(struct extent_buffer *eb)
3298{
3299 int refs;
3300 if (!eb)
3301 return;
3302
3303 refs = atomic_read(&eb->refs);
3304 while (1) {
3305 if ((!test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) && refs <= 3)
3306 || (test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) &&
3307 refs == 1))
3308 break;
3309 if (atomic_try_cmpxchg(&eb->refs, &refs, refs - 1))
3310 return;
3311 }
3312
3313 spin_lock(&eb->refs_lock);
3314 if (atomic_read(&eb->refs) == 2 &&
3315 test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
3316 !extent_buffer_under_io(eb) &&
3317 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3318 atomic_dec(&eb->refs);
3319
3320 /*
3321 * I know this is terrible, but it's temporary until we stop tracking
3322 * the uptodate bits and such for the extent buffers.
3323 */
3324 release_extent_buffer(eb);
3325}
3326
3327void free_extent_buffer_stale(struct extent_buffer *eb)
3328{
3329 if (!eb)
3330 return;
3331
3332 spin_lock(&eb->refs_lock);
3333 set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
3334
3335 if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
3336 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3337 atomic_dec(&eb->refs);
3338 release_extent_buffer(eb);
3339}
3340
3341static void btree_clear_folio_dirty(struct folio *folio)
3342{
3343 ASSERT(folio_test_dirty(folio));
3344 ASSERT(folio_test_locked(folio));
3345 folio_clear_dirty_for_io(folio);
3346 xa_lock_irq(&folio->mapping->i_pages);
3347 if (!folio_test_dirty(folio))
3348 __xa_clear_mark(&folio->mapping->i_pages,
3349 folio_index(folio), PAGECACHE_TAG_DIRTY);
3350 xa_unlock_irq(&folio->mapping->i_pages);
3351}
3352
3353static void clear_subpage_extent_buffer_dirty(const struct extent_buffer *eb)
3354{
3355 struct btrfs_fs_info *fs_info = eb->fs_info;
3356 struct folio *folio = eb->folios[0];
3357 bool last;
3358
3359 /* btree_clear_folio_dirty() needs page locked. */
3360 folio_lock(folio);
3361 last = btrfs_subpage_clear_and_test_dirty(fs_info, folio, eb->start, eb->len);
3362 if (last)
3363 btree_clear_folio_dirty(folio);
3364 folio_unlock(folio);
3365 WARN_ON(atomic_read(&eb->refs) == 0);
3366}
3367
3368void btrfs_clear_buffer_dirty(struct btrfs_trans_handle *trans,
3369 struct extent_buffer *eb)
3370{
3371 struct btrfs_fs_info *fs_info = eb->fs_info;
3372 int num_folios;
3373
3374 btrfs_assert_tree_write_locked(eb);
3375
3376 if (trans && btrfs_header_generation(eb) != trans->transid)
3377 return;
3378
3379 /*
3380 * Instead of clearing the dirty flag off of the buffer, mark it as
3381 * EXTENT_BUFFER_ZONED_ZEROOUT. This allows us to preserve
3382 * write-ordering in zoned mode, without the need to later re-dirty
3383 * the extent_buffer.
3384 *
3385 * The actual zeroout of the buffer will happen later in
3386 * btree_csum_one_bio.
3387 */
3388 if (btrfs_is_zoned(fs_info) && test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3389 set_bit(EXTENT_BUFFER_ZONED_ZEROOUT, &eb->bflags);
3390 return;
3391 }
3392
3393 if (!test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags))
3394 return;
3395
3396 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes, -eb->len,
3397 fs_info->dirty_metadata_batch);
3398
3399 if (eb->fs_info->nodesize < PAGE_SIZE)
3400 return clear_subpage_extent_buffer_dirty(eb);
3401
3402 num_folios = num_extent_folios(eb);
3403 for (int i = 0; i < num_folios; i++) {
3404 struct folio *folio = eb->folios[i];
3405
3406 if (!folio_test_dirty(folio))
3407 continue;
3408 folio_lock(folio);
3409 btree_clear_folio_dirty(folio);
3410 folio_unlock(folio);
3411 }
3412 WARN_ON(atomic_read(&eb->refs) == 0);
3413}
3414
3415void set_extent_buffer_dirty(struct extent_buffer *eb)
3416{
3417 int num_folios;
3418 bool was_dirty;
3419
3420 check_buffer_tree_ref(eb);
3421
3422 was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
3423
3424 num_folios = num_extent_folios(eb);
3425 WARN_ON(atomic_read(&eb->refs) == 0);
3426 WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
3427 WARN_ON(test_bit(EXTENT_BUFFER_ZONED_ZEROOUT, &eb->bflags));
3428
3429 if (!was_dirty) {
3430 bool subpage = eb->fs_info->nodesize < PAGE_SIZE;
3431
3432 /*
3433 * For subpage case, we can have other extent buffers in the
3434 * same page, and in clear_subpage_extent_buffer_dirty() we
3435 * have to clear page dirty without subpage lock held.
3436 * This can cause race where our page gets dirty cleared after
3437 * we just set it.
3438 *
3439 * Thankfully, clear_subpage_extent_buffer_dirty() has locked
3440 * its page for other reasons, we can use page lock to prevent
3441 * the above race.
3442 */
3443 if (subpage)
3444 lock_page(folio_page(eb->folios[0], 0));
3445 for (int i = 0; i < num_folios; i++)
3446 btrfs_folio_set_dirty(eb->fs_info, eb->folios[i],
3447 eb->start, eb->len);
3448 if (subpage)
3449 unlock_page(folio_page(eb->folios[0], 0));
3450 percpu_counter_add_batch(&eb->fs_info->dirty_metadata_bytes,
3451 eb->len,
3452 eb->fs_info->dirty_metadata_batch);
3453 }
3454#ifdef CONFIG_BTRFS_DEBUG
3455 for (int i = 0; i < num_folios; i++)
3456 ASSERT(folio_test_dirty(eb->folios[i]));
3457#endif
3458}
3459
3460void clear_extent_buffer_uptodate(struct extent_buffer *eb)
3461{
3462 struct btrfs_fs_info *fs_info = eb->fs_info;
3463 int num_folios = num_extent_folios(eb);
3464
3465 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3466 for (int i = 0; i < num_folios; i++) {
3467 struct folio *folio = eb->folios[i];
3468
3469 if (!folio)
3470 continue;
3471
3472 /*
3473 * This is special handling for metadata subpage, as regular
3474 * btrfs_is_subpage() can not handle cloned/dummy metadata.
3475 */
3476 if (fs_info->nodesize >= PAGE_SIZE)
3477 folio_clear_uptodate(folio);
3478 else
3479 btrfs_subpage_clear_uptodate(fs_info, folio,
3480 eb->start, eb->len);
3481 }
3482}
3483
3484void set_extent_buffer_uptodate(struct extent_buffer *eb)
3485{
3486 struct btrfs_fs_info *fs_info = eb->fs_info;
3487 int num_folios = num_extent_folios(eb);
3488
3489 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3490 for (int i = 0; i < num_folios; i++) {
3491 struct folio *folio = eb->folios[i];
3492
3493 /*
3494 * This is special handling for metadata subpage, as regular
3495 * btrfs_is_subpage() can not handle cloned/dummy metadata.
3496 */
3497 if (fs_info->nodesize >= PAGE_SIZE)
3498 folio_mark_uptodate(folio);
3499 else
3500 btrfs_subpage_set_uptodate(fs_info, folio,
3501 eb->start, eb->len);
3502 }
3503}
3504
3505static void clear_extent_buffer_reading(struct extent_buffer *eb)
3506{
3507 clear_bit(EXTENT_BUFFER_READING, &eb->bflags);
3508 smp_mb__after_atomic();
3509 wake_up_bit(&eb->bflags, EXTENT_BUFFER_READING);
3510}
3511
3512static void end_bbio_meta_read(struct btrfs_bio *bbio)
3513{
3514 struct extent_buffer *eb = bbio->private;
3515 struct btrfs_fs_info *fs_info = eb->fs_info;
3516 bool uptodate = !bbio->bio.bi_status;
3517 struct folio_iter fi;
3518 u32 bio_offset = 0;
3519
3520 /*
3521 * If the extent buffer is marked UPTODATE before the read operation
3522 * completes, other calls to read_extent_buffer_pages() will return
3523 * early without waiting for the read to finish, causing data races.
3524 */
3525 WARN_ON(test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags));
3526
3527 eb->read_mirror = bbio->mirror_num;
3528
3529 if (uptodate &&
3530 btrfs_validate_extent_buffer(eb, &bbio->parent_check) < 0)
3531 uptodate = false;
3532
3533 if (uptodate) {
3534 set_extent_buffer_uptodate(eb);
3535 } else {
3536 clear_extent_buffer_uptodate(eb);
3537 set_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
3538 }
3539
3540 bio_for_each_folio_all(fi, &bbio->bio) {
3541 struct folio *folio = fi.folio;
3542 u64 start = eb->start + bio_offset;
3543 u32 len = fi.length;
3544
3545 if (uptodate)
3546 btrfs_folio_set_uptodate(fs_info, folio, start, len);
3547 else
3548 btrfs_folio_clear_uptodate(fs_info, folio, start, len);
3549
3550 bio_offset += len;
3551 }
3552
3553 clear_extent_buffer_reading(eb);
3554 free_extent_buffer(eb);
3555
3556 bio_put(&bbio->bio);
3557}
3558
3559int read_extent_buffer_pages(struct extent_buffer *eb, int wait, int mirror_num,
3560 const struct btrfs_tree_parent_check *check)
3561{
3562 struct btrfs_bio *bbio;
3563 bool ret;
3564
3565 if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3566 return 0;
3567
3568 /*
3569 * We could have had EXTENT_BUFFER_UPTODATE cleared by the write
3570 * operation, which could potentially still be in flight. In this case
3571 * we simply want to return an error.
3572 */
3573 if (unlikely(test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)))
3574 return -EIO;
3575
3576 /* Someone else is already reading the buffer, just wait for it. */
3577 if (test_and_set_bit(EXTENT_BUFFER_READING, &eb->bflags))
3578 goto done;
3579
3580 /*
3581 * Between the initial test_bit(EXTENT_BUFFER_UPTODATE) and the above
3582 * test_and_set_bit(EXTENT_BUFFER_READING), someone else could have
3583 * started and finished reading the same eb. In this case, UPTODATE
3584 * will now be set, and we shouldn't read it in again.
3585 */
3586 if (unlikely(test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))) {
3587 clear_extent_buffer_reading(eb);
3588 return 0;
3589 }
3590
3591 clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
3592 eb->read_mirror = 0;
3593 check_buffer_tree_ref(eb);
3594 atomic_inc(&eb->refs);
3595
3596 bbio = btrfs_bio_alloc(INLINE_EXTENT_BUFFER_PAGES,
3597 REQ_OP_READ | REQ_META, eb->fs_info,
3598 end_bbio_meta_read, eb);
3599 bbio->bio.bi_iter.bi_sector = eb->start >> SECTOR_SHIFT;
3600 bbio->inode = BTRFS_I(eb->fs_info->btree_inode);
3601 bbio->file_offset = eb->start;
3602 memcpy(&bbio->parent_check, check, sizeof(*check));
3603 if (eb->fs_info->nodesize < PAGE_SIZE) {
3604 ret = bio_add_folio(&bbio->bio, eb->folios[0], eb->len,
3605 eb->start - folio_pos(eb->folios[0]));
3606 ASSERT(ret);
3607 } else {
3608 int num_folios = num_extent_folios(eb);
3609
3610 for (int i = 0; i < num_folios; i++) {
3611 struct folio *folio = eb->folios[i];
3612
3613 ret = bio_add_folio(&bbio->bio, folio, eb->folio_size, 0);
3614 ASSERT(ret);
3615 }
3616 }
3617 btrfs_submit_bbio(bbio, mirror_num);
3618
3619done:
3620 if (wait == WAIT_COMPLETE) {
3621 wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_READING, TASK_UNINTERRUPTIBLE);
3622 if (!test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3623 return -EIO;
3624 }
3625
3626 return 0;
3627}
3628
3629static bool report_eb_range(const struct extent_buffer *eb, unsigned long start,
3630 unsigned long len)
3631{
3632 btrfs_warn(eb->fs_info,
3633 "access to eb bytenr %llu len %u out of range start %lu len %lu",
3634 eb->start, eb->len, start, len);
3635 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
3636
3637 return true;
3638}
3639
3640/*
3641 * Check if the [start, start + len) range is valid before reading/writing
3642 * the eb.
3643 * NOTE: @start and @len are offset inside the eb, not logical address.
3644 *
3645 * Caller should not touch the dst/src memory if this function returns error.
3646 */
3647static inline int check_eb_range(const struct extent_buffer *eb,
3648 unsigned long start, unsigned long len)
3649{
3650 unsigned long offset;
3651
3652 /* start, start + len should not go beyond eb->len nor overflow */
3653 if (unlikely(check_add_overflow(start, len, &offset) || offset > eb->len))
3654 return report_eb_range(eb, start, len);
3655
3656 return false;
3657}
3658
3659void read_extent_buffer(const struct extent_buffer *eb, void *dstv,
3660 unsigned long start, unsigned long len)
3661{
3662 const int unit_size = eb->folio_size;
3663 size_t cur;
3664 size_t offset;
3665 char *dst = (char *)dstv;
3666 unsigned long i = get_eb_folio_index(eb, start);
3667
3668 if (check_eb_range(eb, start, len)) {
3669 /*
3670 * Invalid range hit, reset the memory, so callers won't get
3671 * some random garbage for their uninitialized memory.
3672 */
3673 memset(dstv, 0, len);
3674 return;
3675 }
3676
3677 if (eb->addr) {
3678 memcpy(dstv, eb->addr + start, len);
3679 return;
3680 }
3681
3682 offset = get_eb_offset_in_folio(eb, start);
3683
3684 while (len > 0) {
3685 char *kaddr;
3686
3687 cur = min(len, unit_size - offset);
3688 kaddr = folio_address(eb->folios[i]);
3689 memcpy(dst, kaddr + offset, cur);
3690
3691 dst += cur;
3692 len -= cur;
3693 offset = 0;
3694 i++;
3695 }
3696}
3697
3698int read_extent_buffer_to_user_nofault(const struct extent_buffer *eb,
3699 void __user *dstv,
3700 unsigned long start, unsigned long len)
3701{
3702 const int unit_size = eb->folio_size;
3703 size_t cur;
3704 size_t offset;
3705 char __user *dst = (char __user *)dstv;
3706 unsigned long i = get_eb_folio_index(eb, start);
3707 int ret = 0;
3708
3709 WARN_ON(start > eb->len);
3710 WARN_ON(start + len > eb->start + eb->len);
3711
3712 if (eb->addr) {
3713 if (copy_to_user_nofault(dstv, eb->addr + start, len))
3714 ret = -EFAULT;
3715 return ret;
3716 }
3717
3718 offset = get_eb_offset_in_folio(eb, start);
3719
3720 while (len > 0) {
3721 char *kaddr;
3722
3723 cur = min(len, unit_size - offset);
3724 kaddr = folio_address(eb->folios[i]);
3725 if (copy_to_user_nofault(dst, kaddr + offset, cur)) {
3726 ret = -EFAULT;
3727 break;
3728 }
3729
3730 dst += cur;
3731 len -= cur;
3732 offset = 0;
3733 i++;
3734 }
3735
3736 return ret;
3737}
3738
3739int memcmp_extent_buffer(const struct extent_buffer *eb, const void *ptrv,
3740 unsigned long start, unsigned long len)
3741{
3742 const int unit_size = eb->folio_size;
3743 size_t cur;
3744 size_t offset;
3745 char *kaddr;
3746 char *ptr = (char *)ptrv;
3747 unsigned long i = get_eb_folio_index(eb, start);
3748 int ret = 0;
3749
3750 if (check_eb_range(eb, start, len))
3751 return -EINVAL;
3752
3753 if (eb->addr)
3754 return memcmp(ptrv, eb->addr + start, len);
3755
3756 offset = get_eb_offset_in_folio(eb, start);
3757
3758 while (len > 0) {
3759 cur = min(len, unit_size - offset);
3760 kaddr = folio_address(eb->folios[i]);
3761 ret = memcmp(ptr, kaddr + offset, cur);
3762 if (ret)
3763 break;
3764
3765 ptr += cur;
3766 len -= cur;
3767 offset = 0;
3768 i++;
3769 }
3770 return ret;
3771}
3772
3773/*
3774 * Check that the extent buffer is uptodate.
3775 *
3776 * For regular sector size == PAGE_SIZE case, check if @page is uptodate.
3777 * For subpage case, check if the range covered by the eb has EXTENT_UPTODATE.
3778 */
3779static void assert_eb_folio_uptodate(const struct extent_buffer *eb, int i)
3780{
3781 struct btrfs_fs_info *fs_info = eb->fs_info;
3782 struct folio *folio = eb->folios[i];
3783
3784 ASSERT(folio);
3785
3786 /*
3787 * If we are using the commit root we could potentially clear a page
3788 * Uptodate while we're using the extent buffer that we've previously
3789 * looked up. We don't want to complain in this case, as the page was
3790 * valid before, we just didn't write it out. Instead we want to catch
3791 * the case where we didn't actually read the block properly, which
3792 * would have !PageUptodate and !EXTENT_BUFFER_WRITE_ERR.
3793 */
3794 if (test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3795 return;
3796
3797 if (fs_info->nodesize < PAGE_SIZE) {
3798 folio = eb->folios[0];
3799 ASSERT(i == 0);
3800 if (WARN_ON(!btrfs_subpage_test_uptodate(fs_info, folio,
3801 eb->start, eb->len)))
3802 btrfs_subpage_dump_bitmap(fs_info, folio, eb->start, eb->len);
3803 } else {
3804 WARN_ON(!folio_test_uptodate(folio));
3805 }
3806}
3807
3808static void __write_extent_buffer(const struct extent_buffer *eb,
3809 const void *srcv, unsigned long start,
3810 unsigned long len, bool use_memmove)
3811{
3812 const int unit_size = eb->folio_size;
3813 size_t cur;
3814 size_t offset;
3815 char *kaddr;
3816 const char *src = (const char *)srcv;
3817 unsigned long i = get_eb_folio_index(eb, start);
3818 /* For unmapped (dummy) ebs, no need to check their uptodate status. */
3819 const bool check_uptodate = !test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
3820
3821 if (check_eb_range(eb, start, len))
3822 return;
3823
3824 if (eb->addr) {
3825 if (use_memmove)
3826 memmove(eb->addr + start, srcv, len);
3827 else
3828 memcpy(eb->addr + start, srcv, len);
3829 return;
3830 }
3831
3832 offset = get_eb_offset_in_folio(eb, start);
3833
3834 while (len > 0) {
3835 if (check_uptodate)
3836 assert_eb_folio_uptodate(eb, i);
3837
3838 cur = min(len, unit_size - offset);
3839 kaddr = folio_address(eb->folios[i]);
3840 if (use_memmove)
3841 memmove(kaddr + offset, src, cur);
3842 else
3843 memcpy(kaddr + offset, src, cur);
3844
3845 src += cur;
3846 len -= cur;
3847 offset = 0;
3848 i++;
3849 }
3850}
3851
3852void write_extent_buffer(const struct extent_buffer *eb, const void *srcv,
3853 unsigned long start, unsigned long len)
3854{
3855 return __write_extent_buffer(eb, srcv, start, len, false);
3856}
3857
3858static void memset_extent_buffer(const struct extent_buffer *eb, int c,
3859 unsigned long start, unsigned long len)
3860{
3861 const int unit_size = eb->folio_size;
3862 unsigned long cur = start;
3863
3864 if (eb->addr) {
3865 memset(eb->addr + start, c, len);
3866 return;
3867 }
3868
3869 while (cur < start + len) {
3870 unsigned long index = get_eb_folio_index(eb, cur);
3871 unsigned int offset = get_eb_offset_in_folio(eb, cur);
3872 unsigned int cur_len = min(start + len - cur, unit_size - offset);
3873
3874 assert_eb_folio_uptodate(eb, index);
3875 memset(folio_address(eb->folios[index]) + offset, c, cur_len);
3876
3877 cur += cur_len;
3878 }
3879}
3880
3881void memzero_extent_buffer(const struct extent_buffer *eb, unsigned long start,
3882 unsigned long len)
3883{
3884 if (check_eb_range(eb, start, len))
3885 return;
3886 return memset_extent_buffer(eb, 0, start, len);
3887}
3888
3889void copy_extent_buffer_full(const struct extent_buffer *dst,
3890 const struct extent_buffer *src)
3891{
3892 const int unit_size = src->folio_size;
3893 unsigned long cur = 0;
3894
3895 ASSERT(dst->len == src->len);
3896
3897 while (cur < src->len) {
3898 unsigned long index = get_eb_folio_index(src, cur);
3899 unsigned long offset = get_eb_offset_in_folio(src, cur);
3900 unsigned long cur_len = min(src->len, unit_size - offset);
3901 void *addr = folio_address(src->folios[index]) + offset;
3902
3903 write_extent_buffer(dst, addr, cur, cur_len);
3904
3905 cur += cur_len;
3906 }
3907}
3908
3909void copy_extent_buffer(const struct extent_buffer *dst,
3910 const struct extent_buffer *src,
3911 unsigned long dst_offset, unsigned long src_offset,
3912 unsigned long len)
3913{
3914 const int unit_size = dst->folio_size;
3915 u64 dst_len = dst->len;
3916 size_t cur;
3917 size_t offset;
3918 char *kaddr;
3919 unsigned long i = get_eb_folio_index(dst, dst_offset);
3920
3921 if (check_eb_range(dst, dst_offset, len) ||
3922 check_eb_range(src, src_offset, len))
3923 return;
3924
3925 WARN_ON(src->len != dst_len);
3926
3927 offset = get_eb_offset_in_folio(dst, dst_offset);
3928
3929 while (len > 0) {
3930 assert_eb_folio_uptodate(dst, i);
3931
3932 cur = min(len, (unsigned long)(unit_size - offset));
3933
3934 kaddr = folio_address(dst->folios[i]);
3935 read_extent_buffer(src, kaddr + offset, src_offset, cur);
3936
3937 src_offset += cur;
3938 len -= cur;
3939 offset = 0;
3940 i++;
3941 }
3942}
3943
3944/*
3945 * Calculate the folio and offset of the byte containing the given bit number.
3946 *
3947 * @eb: the extent buffer
3948 * @start: offset of the bitmap item in the extent buffer
3949 * @nr: bit number
3950 * @folio_index: return index of the folio in the extent buffer that contains
3951 * the given bit number
3952 * @folio_offset: return offset into the folio given by folio_index
3953 *
3954 * This helper hides the ugliness of finding the byte in an extent buffer which
3955 * contains a given bit.
3956 */
3957static inline void eb_bitmap_offset(const struct extent_buffer *eb,
3958 unsigned long start, unsigned long nr,
3959 unsigned long *folio_index,
3960 size_t *folio_offset)
3961{
3962 size_t byte_offset = BIT_BYTE(nr);
3963 size_t offset;
3964
3965 /*
3966 * The byte we want is the offset of the extent buffer + the offset of
3967 * the bitmap item in the extent buffer + the offset of the byte in the
3968 * bitmap item.
3969 */
3970 offset = start + offset_in_eb_folio(eb, eb->start) + byte_offset;
3971
3972 *folio_index = offset >> eb->folio_shift;
3973 *folio_offset = offset_in_eb_folio(eb, offset);
3974}
3975
3976/*
3977 * Determine whether a bit in a bitmap item is set.
3978 *
3979 * @eb: the extent buffer
3980 * @start: offset of the bitmap item in the extent buffer
3981 * @nr: bit number to test
3982 */
3983int extent_buffer_test_bit(const struct extent_buffer *eb, unsigned long start,
3984 unsigned long nr)
3985{
3986 unsigned long i;
3987 size_t offset;
3988 u8 *kaddr;
3989
3990 eb_bitmap_offset(eb, start, nr, &i, &offset);
3991 assert_eb_folio_uptodate(eb, i);
3992 kaddr = folio_address(eb->folios[i]);
3993 return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
3994}
3995
3996static u8 *extent_buffer_get_byte(const struct extent_buffer *eb, unsigned long bytenr)
3997{
3998 unsigned long index = get_eb_folio_index(eb, bytenr);
3999
4000 if (check_eb_range(eb, bytenr, 1))
4001 return NULL;
4002 return folio_address(eb->folios[index]) + get_eb_offset_in_folio(eb, bytenr);
4003}
4004
4005/*
4006 * Set an area of a bitmap to 1.
4007 *
4008 * @eb: the extent buffer
4009 * @start: offset of the bitmap item in the extent buffer
4010 * @pos: bit number of the first bit
4011 * @len: number of bits to set
4012 */
4013void extent_buffer_bitmap_set(const struct extent_buffer *eb, unsigned long start,
4014 unsigned long pos, unsigned long len)
4015{
4016 unsigned int first_byte = start + BIT_BYTE(pos);
4017 unsigned int last_byte = start + BIT_BYTE(pos + len - 1);
4018 const bool same_byte = (first_byte == last_byte);
4019 u8 mask = BITMAP_FIRST_BYTE_MASK(pos);
4020 u8 *kaddr;
4021
4022 if (same_byte)
4023 mask &= BITMAP_LAST_BYTE_MASK(pos + len);
4024
4025 /* Handle the first byte. */
4026 kaddr = extent_buffer_get_byte(eb, first_byte);
4027 *kaddr |= mask;
4028 if (same_byte)
4029 return;
4030
4031 /* Handle the byte aligned part. */
4032 ASSERT(first_byte + 1 <= last_byte);
4033 memset_extent_buffer(eb, 0xff, first_byte + 1, last_byte - first_byte - 1);
4034
4035 /* Handle the last byte. */
4036 kaddr = extent_buffer_get_byte(eb, last_byte);
4037 *kaddr |= BITMAP_LAST_BYTE_MASK(pos + len);
4038}
4039
4040
4041/*
4042 * Clear an area of a bitmap.
4043 *
4044 * @eb: the extent buffer
4045 * @start: offset of the bitmap item in the extent buffer
4046 * @pos: bit number of the first bit
4047 * @len: number of bits to clear
4048 */
4049void extent_buffer_bitmap_clear(const struct extent_buffer *eb,
4050 unsigned long start, unsigned long pos,
4051 unsigned long len)
4052{
4053 unsigned int first_byte = start + BIT_BYTE(pos);
4054 unsigned int last_byte = start + BIT_BYTE(pos + len - 1);
4055 const bool same_byte = (first_byte == last_byte);
4056 u8 mask = BITMAP_FIRST_BYTE_MASK(pos);
4057 u8 *kaddr;
4058
4059 if (same_byte)
4060 mask &= BITMAP_LAST_BYTE_MASK(pos + len);
4061
4062 /* Handle the first byte. */
4063 kaddr = extent_buffer_get_byte(eb, first_byte);
4064 *kaddr &= ~mask;
4065 if (same_byte)
4066 return;
4067
4068 /* Handle the byte aligned part. */
4069 ASSERT(first_byte + 1 <= last_byte);
4070 memset_extent_buffer(eb, 0, first_byte + 1, last_byte - first_byte - 1);
4071
4072 /* Handle the last byte. */
4073 kaddr = extent_buffer_get_byte(eb, last_byte);
4074 *kaddr &= ~BITMAP_LAST_BYTE_MASK(pos + len);
4075}
4076
4077static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
4078{
4079 unsigned long distance = (src > dst) ? src - dst : dst - src;
4080 return distance < len;
4081}
4082
4083void memcpy_extent_buffer(const struct extent_buffer *dst,
4084 unsigned long dst_offset, unsigned long src_offset,
4085 unsigned long len)
4086{
4087 const int unit_size = dst->folio_size;
4088 unsigned long cur_off = 0;
4089
4090 if (check_eb_range(dst, dst_offset, len) ||
4091 check_eb_range(dst, src_offset, len))
4092 return;
4093
4094 if (dst->addr) {
4095 const bool use_memmove = areas_overlap(src_offset, dst_offset, len);
4096
4097 if (use_memmove)
4098 memmove(dst->addr + dst_offset, dst->addr + src_offset, len);
4099 else
4100 memcpy(dst->addr + dst_offset, dst->addr + src_offset, len);
4101 return;
4102 }
4103
4104 while (cur_off < len) {
4105 unsigned long cur_src = cur_off + src_offset;
4106 unsigned long folio_index = get_eb_folio_index(dst, cur_src);
4107 unsigned long folio_off = get_eb_offset_in_folio(dst, cur_src);
4108 unsigned long cur_len = min(src_offset + len - cur_src,
4109 unit_size - folio_off);
4110 void *src_addr = folio_address(dst->folios[folio_index]) + folio_off;
4111 const bool use_memmove = areas_overlap(src_offset + cur_off,
4112 dst_offset + cur_off, cur_len);
4113
4114 __write_extent_buffer(dst, src_addr, dst_offset + cur_off, cur_len,
4115 use_memmove);
4116 cur_off += cur_len;
4117 }
4118}
4119
4120void memmove_extent_buffer(const struct extent_buffer *dst,
4121 unsigned long dst_offset, unsigned long src_offset,
4122 unsigned long len)
4123{
4124 unsigned long dst_end = dst_offset + len - 1;
4125 unsigned long src_end = src_offset + len - 1;
4126
4127 if (check_eb_range(dst, dst_offset, len) ||
4128 check_eb_range(dst, src_offset, len))
4129 return;
4130
4131 if (dst_offset < src_offset) {
4132 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
4133 return;
4134 }
4135
4136 if (dst->addr) {
4137 memmove(dst->addr + dst_offset, dst->addr + src_offset, len);
4138 return;
4139 }
4140
4141 while (len > 0) {
4142 unsigned long src_i;
4143 size_t cur;
4144 size_t dst_off_in_folio;
4145 size_t src_off_in_folio;
4146 void *src_addr;
4147 bool use_memmove;
4148
4149 src_i = get_eb_folio_index(dst, src_end);
4150
4151 dst_off_in_folio = get_eb_offset_in_folio(dst, dst_end);
4152 src_off_in_folio = get_eb_offset_in_folio(dst, src_end);
4153
4154 cur = min_t(unsigned long, len, src_off_in_folio + 1);
4155 cur = min(cur, dst_off_in_folio + 1);
4156
4157 src_addr = folio_address(dst->folios[src_i]) + src_off_in_folio -
4158 cur + 1;
4159 use_memmove = areas_overlap(src_end - cur + 1, dst_end - cur + 1,
4160 cur);
4161
4162 __write_extent_buffer(dst, src_addr, dst_end - cur + 1, cur,
4163 use_memmove);
4164
4165 dst_end -= cur;
4166 src_end -= cur;
4167 len -= cur;
4168 }
4169}
4170
4171#define GANG_LOOKUP_SIZE 16
4172static struct extent_buffer *get_next_extent_buffer(
4173 const struct btrfs_fs_info *fs_info, struct folio *folio, u64 bytenr)
4174{
4175 struct extent_buffer *gang[GANG_LOOKUP_SIZE];
4176 struct extent_buffer *found = NULL;
4177 u64 folio_start = folio_pos(folio);
4178 u64 cur = folio_start;
4179
4180 ASSERT(in_range(bytenr, folio_start, PAGE_SIZE));
4181 lockdep_assert_held(&fs_info->buffer_lock);
4182
4183 while (cur < folio_start + PAGE_SIZE) {
4184 int ret;
4185 int i;
4186
4187 ret = radix_tree_gang_lookup(&fs_info->buffer_radix,
4188 (void **)gang, cur >> fs_info->sectorsize_bits,
4189 min_t(unsigned int, GANG_LOOKUP_SIZE,
4190 PAGE_SIZE / fs_info->nodesize));
4191 if (ret == 0)
4192 goto out;
4193 for (i = 0; i < ret; i++) {
4194 /* Already beyond page end */
4195 if (gang[i]->start >= folio_start + PAGE_SIZE)
4196 goto out;
4197 /* Found one */
4198 if (gang[i]->start >= bytenr) {
4199 found = gang[i];
4200 goto out;
4201 }
4202 }
4203 cur = gang[ret - 1]->start + gang[ret - 1]->len;
4204 }
4205out:
4206 return found;
4207}
4208
4209static int try_release_subpage_extent_buffer(struct folio *folio)
4210{
4211 struct btrfs_fs_info *fs_info = folio_to_fs_info(folio);
4212 u64 cur = folio_pos(folio);
4213 const u64 end = cur + PAGE_SIZE;
4214 int ret;
4215
4216 while (cur < end) {
4217 struct extent_buffer *eb = NULL;
4218
4219 /*
4220 * Unlike try_release_extent_buffer() which uses folio private
4221 * to grab buffer, for subpage case we rely on radix tree, thus
4222 * we need to ensure radix tree consistency.
4223 *
4224 * We also want an atomic snapshot of the radix tree, thus go
4225 * with spinlock rather than RCU.
4226 */
4227 spin_lock(&fs_info->buffer_lock);
4228 eb = get_next_extent_buffer(fs_info, folio, cur);
4229 if (!eb) {
4230 /* No more eb in the page range after or at cur */
4231 spin_unlock(&fs_info->buffer_lock);
4232 break;
4233 }
4234 cur = eb->start + eb->len;
4235
4236 /*
4237 * The same as try_release_extent_buffer(), to ensure the eb
4238 * won't disappear out from under us.
4239 */
4240 spin_lock(&eb->refs_lock);
4241 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
4242 spin_unlock(&eb->refs_lock);
4243 spin_unlock(&fs_info->buffer_lock);
4244 break;
4245 }
4246 spin_unlock(&fs_info->buffer_lock);
4247
4248 /*
4249 * If tree ref isn't set then we know the ref on this eb is a
4250 * real ref, so just return, this eb will likely be freed soon
4251 * anyway.
4252 */
4253 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4254 spin_unlock(&eb->refs_lock);
4255 break;
4256 }
4257
4258 /*
4259 * Here we don't care about the return value, we will always
4260 * check the folio private at the end. And
4261 * release_extent_buffer() will release the refs_lock.
4262 */
4263 release_extent_buffer(eb);
4264 }
4265 /*
4266 * Finally to check if we have cleared folio private, as if we have
4267 * released all ebs in the page, the folio private should be cleared now.
4268 */
4269 spin_lock(&folio->mapping->i_private_lock);
4270 if (!folio_test_private(folio))
4271 ret = 1;
4272 else
4273 ret = 0;
4274 spin_unlock(&folio->mapping->i_private_lock);
4275 return ret;
4276
4277}
4278
4279int try_release_extent_buffer(struct folio *folio)
4280{
4281 struct extent_buffer *eb;
4282
4283 if (folio_to_fs_info(folio)->nodesize < PAGE_SIZE)
4284 return try_release_subpage_extent_buffer(folio);
4285
4286 /*
4287 * We need to make sure nobody is changing folio private, as we rely on
4288 * folio private as the pointer to extent buffer.
4289 */
4290 spin_lock(&folio->mapping->i_private_lock);
4291 if (!folio_test_private(folio)) {
4292 spin_unlock(&folio->mapping->i_private_lock);
4293 return 1;
4294 }
4295
4296 eb = folio_get_private(folio);
4297 BUG_ON(!eb);
4298
4299 /*
4300 * This is a little awful but should be ok, we need to make sure that
4301 * the eb doesn't disappear out from under us while we're looking at
4302 * this page.
4303 */
4304 spin_lock(&eb->refs_lock);
4305 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
4306 spin_unlock(&eb->refs_lock);
4307 spin_unlock(&folio->mapping->i_private_lock);
4308 return 0;
4309 }
4310 spin_unlock(&folio->mapping->i_private_lock);
4311
4312 /*
4313 * If tree ref isn't set then we know the ref on this eb is a real ref,
4314 * so just return, this page will likely be freed soon anyway.
4315 */
4316 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4317 spin_unlock(&eb->refs_lock);
4318 return 0;
4319 }
4320
4321 return release_extent_buffer(eb);
4322}
4323
4324/*
4325 * Attempt to readahead a child block.
4326 *
4327 * @fs_info: the fs_info
4328 * @bytenr: bytenr to read
4329 * @owner_root: objectid of the root that owns this eb
4330 * @gen: generation for the uptodate check, can be 0
4331 * @level: level for the eb
4332 *
4333 * Attempt to readahead a tree block at @bytenr. If @gen is 0 then we do a
4334 * normal uptodate check of the eb, without checking the generation. If we have
4335 * to read the block we will not block on anything.
4336 */
4337void btrfs_readahead_tree_block(struct btrfs_fs_info *fs_info,
4338 u64 bytenr, u64 owner_root, u64 gen, int level)
4339{
4340 struct btrfs_tree_parent_check check = {
4341 .level = level,
4342 .transid = gen
4343 };
4344 struct extent_buffer *eb;
4345 int ret;
4346
4347 eb = btrfs_find_create_tree_block(fs_info, bytenr, owner_root, level);
4348 if (IS_ERR(eb))
4349 return;
4350
4351 if (btrfs_buffer_uptodate(eb, gen, 1)) {
4352 free_extent_buffer(eb);
4353 return;
4354 }
4355
4356 ret = read_extent_buffer_pages(eb, WAIT_NONE, 0, &check);
4357 if (ret < 0)
4358 free_extent_buffer_stale(eb);
4359 else
4360 free_extent_buffer(eb);
4361}
4362
4363/*
4364 * Readahead a node's child block.
4365 *
4366 * @node: parent node we're reading from
4367 * @slot: slot in the parent node for the child we want to read
4368 *
4369 * A helper for btrfs_readahead_tree_block, we simply read the bytenr pointed at
4370 * the slot in the node provided.
4371 */
4372void btrfs_readahead_node_child(struct extent_buffer *node, int slot)
4373{
4374 btrfs_readahead_tree_block(node->fs_info,
4375 btrfs_node_blockptr(node, slot),
4376 btrfs_header_owner(node),
4377 btrfs_node_ptr_generation(node, slot),
4378 btrfs_header_level(node) - 1);
4379}